Let us imagine an alternative history of AI. In this world match all the astonishing progress up to and including 2017, but once the transformer architecture was reported in the seminal “Attention is all you need” paper, there was a small but significant divergence. Once Ilya Sutsekever “jumped out of his chair” on reading the paper, then instead of instructing his team to drop everything and work on just this idea … history went differently.

In this new world, Ilya sat down again and wrote a brilliantly compelling note to all his colleagues at OpenAI and copied the research leads of all other labs. He explained the potential that he saw in this new architecture, but also his fears of the results getting out of hand if it were unleashed on the training data available in human language.

Instead, he proposed an experimental setup that all the labs could restrict themselves to. Once they read his note - improbably enough - they all agreed that this was the way forward, and took it.

ChimpGPT

Instead of unleashing the new architecture on training sets that amount to the entirety of human language, and diving straight into technical challenges of the engineering needed to feed these through the bevy of transformers, the proposal was to conduct a more limited trial. That is, we decided to do another species’ language first.

We could do snails. But it’s not clear how much they talk. We could do dolphins, but - with apologies to Douglas Adams - it’s not clear how much they do beyond mucking about in the water having a good time. So, we’ll choose chimpanzees.

That is, we take all the communications research that exists on chimpanzees, on their facial expressions, body language and sounds, and we put together a machine codable transcript of everything we have ever seen chimpanzees “say” to one another, and what was said in response.

Those who have seen chimps “talk” to one another in troupes can have little doubt that there is fairly sophisticated communication going on. We continue until the dataset is large enough for the transformers to start to bite. These GPT architectures are bewilderingly data-hungry, so this might have run into practical problems. We will certainly need orders of magnitude more chimp data than exists at present. Potentially, we would even need more chimps than exist in the world at present. We’ll breed them, we’ll wait.

And then we train our LLMs on this dataset instead. What might we expect?

On the assumption we are not going to actually do this experiment (we still could do it in reality, of course, we just need enough chimpanzee observations) we are left to imagine, and to draw from what has happened with human language.

It might look something like this. Let us assume that - as needed - the model is given some kind of chimpanzee body - a sufficiently strong and robust robot that can make the exaggerated facial expressions, gestures, and noises that seem essential to chimpanzee communication, and we set this “avatar” to interact within a chimpanzee troupe. In the first few iterations, the chimpanzees would be able to interact with the ChimpGPT model, and it would give some impression of understanding and playing a role, though no doubt it would be quickly rejected by the chimps at first.

Over subsequent releases, and refinements we can hope it would become more and more adept at understanding context and the social dynamics of the chimp society it was dropped into, until - at the frontier - this avatar is able to “converse” as if it were familiar with the forest environment around it, to respond as a chimpanzee would, and to play whatever role we asked of it. True, even a sophisticated model would get things wrong frequently - perhaps it would discount or overweight the presence of predators in a way that made some of the older, greyer chimps laugh knowingly (though I believe they spend a lot of time laughing knowingly in any case). Or these models would reveal that they didn’t know too much about the practicalities of branch strength and how practical it was to move between trees, with hilarious results and perhaps the chimps would get hours of fun by inducing the model to “say” ridiculous things or make amusing errors in social hierarchy, and induce hours of cackling at its mistakes. But, as we fed tuned the models to chimp feedback, and with more communications data, they would get better. And with more and more data, they would - very quickly - get astonishingly good.

In fact, if we were studying this, we might start drawing ourselves some charts for Pan troglodytes, which might look something like this:

Now, where would we expect the capability of these chimp-trained models to top out? I should be clear here. Without the experience of the last few years, I would have expected that the models would manage - at best - the communication and reasoning ability of a pretty sub-par chimpanzee. And I would find that impressive enough.

So it’s with quite some surprise that we can look at some of the intriguing mathematical results from models in our our real world, which are somewhere near the bleeding edge of the reasoning capabilities of the very best thinkers of the field - with the addition of a huge memory, and not limited by human constraints such as fatigue. With the addition of this knowledge, we might predict that a sufficiently well-funded research project the reasoning could go at least to, and perhaps a bit beyond what the very smartest chimps were able to do. Using these social interaction, planning and plotting at the very limits we might predict that the computer model would be able to place themselves as the leader of almost any group they were in. Perhaps using twigs in new and exciting ways to extract insect larvae, extending this to other species - bird’s eggs and so on, and coming up with ways in which to be socially useful, or even dominant in their troupe.

From there, we continue to feed more, and more and more chimp data into larger models. We should expect the charts to continue. To complete tasks at a level of sophistication and effort no chimp ever has - ones that would take weeks, months or even years of sustained chimpanzee mental effort. Ones that no chimpanzee had ever done, simply because they’d be somewhere around the thousands-of-chimp-hours mark on the horizontal axis above, and no chimp had ever been able to concentrate for that long. Perhaps these tasks would be beyond anything any individual chimp had achieved. But where would they go on from there?

Of course we don’t know. But, to put the question bluntly - using chimp input alone - could we expect our models to derive, from the deep patterns in chimp communication, proper “SuperPan” capabilities and come up with - let’s say - renormalisable quantum electrodynamics?

Degrees of shock and awe

My suggestion is that any result like this would be an enormous surprise - and a surprise as much about chimpanzees as about the models. Far larger than the surprise that the Pan models would be able to mimic the behaviour of the most sophisticated chimpanzees (and even a fair amount beyond). What it would imply is that embedded somewhere in the communications of these chimpanzees were concepts and ideas that would allow entirely new levels of abstraction, reasoning, and advance, all lying latent and unused in their communication, but present and evident enough for the models to pick up on them. One might think of it as a kind of extended Turing principle for language - but here applied to a system of facial expressions, grunts and gestures.

Instead, I think we should expect the models to level out. To be able to model and exhibit more and more sophisticated chimp behaviour until they extend beyond what can be achieved by even the most intelligent chimpanzee. Perhaps enough to charm, control or manipulate just about every chimp that was exposed to one of the best avatars. Enough to become leaders of every troupe they (or their avatars) were dropped into.

But if it were trained only on the communications used by chimps, more and more data, more and more compute, would not replicate the cultural and intellectual achievements of humans. These things are - I suggest - simply not derivable from deep learning applied to the patterns in chimpanzee communication and use. Though of course in the real world, we are starting to see that they are derivable from human language and use.

So, what have we learned about superintelligence?

It’s pretty pedestrian to point out the inference intended from our intuition pump to putative limits on the capabilities of current approaches to LLM training to achieve “superintelligence”.1

So perhaps it’s better to fast-forward through this, and instead imagine some moves that we might suggest to allow our imagined ChimpGPT models to break through a “uber-smart-chimpanzee” limit.

Three strategies are fairly clear:

1. Allow models to “talk” to themselves: in some way similar to the reasoning and agentic models that have led to a jump forward in reasoning ability compared to non-reasoning “one-shot” language models. And this could be powerful, particularly this kind of internal “conversation” is missing, or supressed in chimps.2 If so, this strategy might open up reasoning abilities beyond - perhaps even far beyond - those in chimpanzees themselves.

   The issue with this hypothesis is that we have added a qualitative capability that humans have, but chimpanzees do not. In the inference to human language, what is this extra capability? Perhaps there is one, but it would take a much greater understanding of superhuman intelligence than we have, in order to seek it out.

2. Allow several models to reason “against” one another, and improve one another. The analogy here is to what happens in the reinforcement learning approaches of DeepMind’s AlphaGo, where the training approach quickly left large databases of human-derived moves behind, and models trained against one another, moving way beyond human performance. If several such chimp-models worked against one another to try to become a leader of a troupe - or even to find more food than the other for the troupe, they would likewise be able to become orders of magnitude better at these jobs than the original chimps doing these roles - perhaps evolving new ways of communicating on top of the chimpanzee language they started with; their modes of communication becoming ever-more sophisticated alongside their own behaviour.

   I think it’s plausible that by creating such an artificial community, one might indeed move beyond a “smart-chimp” limit. But again, it would be an astonishing shock - one hugely beyond the things we have seen thus far - if this approach made the leap out towards human levels of sophistication. We might well get super-chimp fruit gatherers. But hard to see the path from here through technology, science, abstraction, theories and creating ones own goals in order to break through to human-like levels of sophistication.

3. Move away from chimp communication altogether - and go for a “world model”. That is, to train on a far wider dataset than just chimpanzee communication. On the physics, and biology of the surrounding forest. Of how moving through trees, interacting with light, shade and objects works directly, not simply through the limited communication channels set up by chimpanzees. After all, this is presumably how we ourselves - humans - did manage to break through to different levels of reasoning and intelligence from a place in our evolutionary past in which we - presumably - were communicating in a way quite similarly to present-day chimpanzees.

   Of course this approach is one that is possible, and is one being taken by several (mostly rather early stage and stealthy-but-highly-funded) research groups, as a way to move beyond LLMs. But this is precisely the point - it is not an LLM approach. Indeed it takes arguments similar to, and in addition to the intuition pump given here, and moves away from LLMs as a dead-end. And quite frankly, the deep learning approach has shown relatively limited success on learning directly from real-world physics and surroundings, compared to its quite dizzying approach learning from language. The amount of cash invested in Anthropic - who are not shy about their LLM-or-bust approach versus - say - Yann LeCunn’s start-up is a ratio of around 130:1, and gives you an idea of where the proven progress is so far.

We will not have long to wait. Current LLMs are already approaching the levels where this intuition pump suggests the best approaches will start to level off. If this kind of imagination is right, models will continue to improve marginally, but it will be hard to ignore the diminishing return effects from here on.

Go to www.nih.gov - the National Institutes of Health in the United States, perhaps the single biggest funder of research in the world, and type a random word of your choice into their search bar.

Instantly I am given all the information I need, including the fascinating fact that Bitter Orange has not been found beneficial or damaging to health.

Now try “equity”.

Now try “racism”.

Now try “gender”.

At the same time, the American Society for Microbiology have taken to silently, and retrospectively, editing articles on their website against some a similar list of “banned" words. Authors are not informed at all, but if your article had any of a set of certain words in it, the AMS has silently, and retrospectively, removed them.

It’s not at all clear how many articles have been edited, and how many removed altogether. Some articles have now had their changes reverted, apparently as people start noticing and making noise. The ASM’s statement in response just lies, saying there has been no editing or removal of “any research articles, scientific papers or peer reviewed articles on our website”. (Presumably this squeaks through by using some definition of “research article” approximating to “one we didn’t edit”). It is also - this is the most polite I can be - the single most whiny, condescending, piece of self-justification have ever read.

The National Sciences Foundation have been given a much longer list of words that should be “flagged” if they appear anywhere in the abstract of a study for their approval. It is - presumably unintentionally - hilarious.

So here’s another game.

Try to write a full scientific paper while avoiding using the words “prejudice”, “systemic”, “excluded”, “bias”, “equality” or “marginalize”. Most statistics methods papers will fail. (Even if they do manage to avoid using the word “female”, “females” or “women”.) It is very difficult to write anything scientific without using half the terms scientists employ to talk about methods of avoiding error.

So, the good news: the people trying to implement all this are evidently morons.

The bad news: so far, no-one seems to be standing up to the morons

I’m not entirely sure what I expected as the “extreme free speech” and “anti-censorship” lot came into power in the United States, but I can’t say this was it. If I was expecting anything at all, it was that a country that claims it was founded on liberty, and with a strong Constitution with all those famed checks and balances, would not have swung this far in just a matter of days. And with apparently very little pushback.

There are many ways of seeing this, but one of the clearer is to simply list the worst months for ambulance response times, in order of longest average waits, and see when they fell.

Looking at the top 20, only three 2024-25-season months appear (in red), whereas most are from the 2022-3 season (in blue)

In Wales, comparable data has has taken longer to come through, but it’s here now. And despite similar demand pressures (again, little COVID, bad flu), it’s clear a similar conclusion cannot be made. Using their (slightly different) measures of median times, and their (slightly different) “Amber” category, we see this:

This winter - 2024-5 - just failed to take the top-spot, but it’s claimed many of the others. The horror has been sustained. Getting an ambulance in Wales during the winter of 2024-5 was pretty much impossible.1

As with England, the category that is particularly revealing for Wales is this “second-most-urgent” category - the one you are put in when you are not just about to die, but when a medical response is clearly urgent to prevent serious harm.

For Wales this is their “Amber” Category, and includes indicators like intense chest pain, and indicators of a possible stroke. The Amber category does not have a specific target for response times - unlike the Category 2 for England - but their performance indicators are the administration of care bundles for cardiac and stroke - simply because these are the levels of emergency that the category is designed for.

Given this intent, the median response times to Welsh Amber calls, by ambulance board are scarcely believable. To re-iterate: the vertical axis is a median measure, and gives numbers not in minutes, but in hours.2

It is clear that for the last few years, the very worst place in Britain to have a heart attack is in Swansea, and whatever you do, you need to avoid having it in winter. Especially last month. You can can see the system spiralling increasingly out of control.

Raw desperation

Another indirect but revealing indicator of just how much the whole Welsh system is struggling is the proportion of all calls coded to the most urgent category over time. Red is supposed to be reserved for true life-threatening calls that have to be responded to within eight minutes: not breathing, no pulse, bleeding out.

Since the standards have not changed over this time period, there a variety of ways of reading this chart. It’s possible that people genuinely are more seriously ill at these times, or people avoid calling ambulances for less serious complaints when they know the system is under pressure (though there is no obvious fall-off in total call-out volumes at these times)

But my interpretation would be the most obvious - either the the dispatchers, or the callers themselves, are aware that they will not get any ambulance within hours if they interpret the guidelines or the descriptions of the complaint as usual, and are “upgrading” urgency of calls that are not really life-threatening. We also saw this in England in winter 2022, but it is more dramatic in Wales.

And this of course means that the system has completely broken down. If you are not able to “clear the board” for those who are genuinely in the most life-threatening situation, because the system is so under strain, it is forced to clear the board for everyone, you will certainly miss some.

It is possible that all this is very well known within Wales, and simply not making the national news. But I have not seen it much remarked that there seems a fundamental, and life-threatening, difference between Welsh emergency response and that in the rest of the country.

The CDC host and maintain major datasets, at the federal level, of US data on the progress and history of diseases. But if you try data.cdc.gov right now, you will either get nothing, or this landing page

[Update: 17:58 1/2/2025 - this landing page has been removed; the datasets highlighted below remain offline]

Some links to some databases still work. So - as at the time of writing - do their recommendations on vaccination schedules. Others are not available. In general, the pattern seems to be any data that makes reference to gender or to HIV is more likely to be unavailable. The following appear to be completely down.

1. CDC Atlas of disease

2. CDC Youth Risk Behavior Surveillance System

3. TargetHIV the The CDC/HRSA programme for caring for people with HIV

4. CDC Social Vulnerability Index

(Usual links above, so do check how they’re doing. Not a single one of these works for me as of the afternoon of 1 February 2025.)

This is, of course insane, though - it must be said - entirely in line with what the American people voted for last November. Moves like this were promised in the Project 2025 manifesto.

And the CDC is of course not alone in this: pages relevant to PEPFAR (the President’s fund for the prevention of AIDS), and government databases on international tracking and prediction of famine, have also been removed. But then many other datasets are still there (e.g., I can’t see any obvious missing CDC COVID datasets.)

So, the promise on the landing page may be honoured, and it’s all just temporary while they comply with a couple of Presidential orders1 … but since both the orders themselves, and the CDC’s response of pulling everything is mad … and on the assumption that this madness may continue, many people have been trying to save historical records of these major CDC databases.

Archiving historical information itself is of course no substitute for having regularly updated and cleaned live data for any new development (bird flu is the most obvious one to be concerned about), but having some archive offline is better than the data being lost altogether - either by mistake, or deliberately.

So here are a few resources that people may find helpful:

• All US government pages as they were at the end of the Biden administration (N.B., pages only, not databases) https://eotarchive.org/

• Many of the big tables and CSVs from the CDC datasets. Unclear to me how much is included, but a great deal of valuable tables are clearly there. https://archive.org/details/20250128-cdc-datasets

• A CDC FTP site, which - again at the time of writing - seems to be still up and to have a fair amount of raw data still accessible.

  https://ftp.cdc.gov/

• The CDC’s geospatial Social Vulnerability Index layers have been extracted and are now on the Esri Living Atlas: livingatlas.arcgis.com

• A progress report on a comprehensive torrent file being put together by the reddit community: r/Datahoarder. https://www.reddit.com/r/DataHoarder/comments/1ibnjbb/comment/m9l4ajh/

• A tracker for journalists or researchers who have downloaded further copies of CDC data - only use if you are comfortable being identified - it may help to know who has what. https://docs.google.com/forms/d/e/1FAIpQLScdi6gcBKKqoK3iV4uHsGFnaSlgT2vAfHaCr_6SdNZJkYWRhQ/viewform

There are no doubt more, quieter approaches to keeping this data out from political filters and possible deletion, and I will not talk about those who would rather keep things quiet, but please let me know if there are other public efforts that should be linked here.

Because so far, the American public’s response to all this - and wider moves - looks to me to be total apathy. And I would rather like to be proved wrong about this.

Hospitals:

• Derriford

• Treliske

• 2 x Hampshire (Basingstoke, Royal H)

• 4 x Birmingham (Good Hope, Heartlands, Solihull, QEH)

• Northampton

• Kettering

• Royal Liverpool UH

• Wirral

• Warwick

• Cambridge UH

Ambulance Services:

• East Midlands

• Wales

There are probably more.

And December’s ambulance times are heading well back to the bad old days (though not as bad as I predicted - I think this is because the really bad times covered only the last week or so of December, and were “averaged out” by the rest of the month. However you cut it, Late 2022 was still significantly worse.1

Now, almost all of this pressure had been triggered by a nasty season of flu (with COVID hardly featuring, RSV peaking too prematurely, and adding bit of norovirus as the world’s least favourite minor player.)

And the flu … I thought - from some data released on Monday - might have peaked. Might have. So I wrote it up with some suggestive charts and extra hope: Thank god? NHS edition.

Today, NHS sitrep numbers came out - the beds occupied gives the most direct measure of flu pressure on hospitals. And…

We can remove that question-mark. Flu pressure has peaked.

(So has noro - and at about the same time - no doubt to every NHS professional’s extreme relief).

It’s not over of course. The system still can’t cope even at these levels of seasonal virus, and COVID is creeping back up to give a bit of a sting in the tail too. But the major factor driving the desperation this winter has peaked; we can look forward it it getting better from here.

So, when we look for reasons for the the terrible wait times and ambulance service “incident” declarations (focused in Wales and the Midlands - and see passim for predictions on this), the situation is fairly simple. The fundamental reason is structural weakness for capacity surges in the NHS. But the trigger is almost entirely down to a bad (but not in any way unprecedented) season of flu. Influenza A to be precise.

So, looking at previous years, we can predict that December’s ambulance times (reported on Thursday 9th) will be pretty diabolical. Perhaps pushing an hour for England average, Wales worse. Most likely the second-or-third-worst month of all time - though they will remain a good way behind December 2022, when flu was higher and had a COVID wave on top

But this is more nowcasting than forecasting. The data isn’t out yet, but these times have already been experienced by patients. This morning (7th January) East Midlands ambulance service just declared their first ever critical incident, and there are reports of are 20+ hour waits in A&E in hospitals in the South West. Is this going to get better or worse - can we actually forecast the future?

Yes, I think we can. Since this peak of pressure is fairly simply driven by flu, we can look to those trends, and trust (with fingers firmly crossed) that it follows a simple “peak and fall” pattern.1

So, looking at the NOIDS causative agent reports (lab tests of samples) we can see that - at least in absolute test numbers - we are just about peaking. Thank god

But let’s also be cautious. Three big reasons for this:

1. There are absolute numbers, and this of course depends on the number of tests run. And we are talking about a holiday period, so perhaps the apparent peak simply reflects a lack of testing, and it’ll be back up next week.2

2. Flu doesn’t always give a single, neat peak. Last year shows a dramatic - though low-level - example of this, where the true peak came in early February, after previously appearing to fall. Rates could surge again.

3. COVID could come and do an “handover”. It’s rising now (from a low base, and probably too late to do a proper “double-demic”) but if it was dramatic, it could hand over straight from one respiratory peak to another, and give the result of constant, or even increasing pressure.

I’m optimistic about all three of these - I don’t think any of them are a big enough effect to detract from the main peak-and-fall pattern. But I cannot prove that any of them are invalid - we should keep them in mind as we look at the next few days.

So, notwithstanding these three caveats, I think we should expect that pressure on hospitals will drag a few days behind the patterns in the fundamental triggering mechanism (because illnesses, hospital stays and bed capacity take several days to work through the system).

So we are probably looking at just about the height of pressure at the moment, and can - with fingers firmly crossed - expect it to lift in coming days, and weeks. It is unlikely to be as dramatic as the “January miracle” of 2023 - you can see that the peak is not as dramatically up and down as it was there. But I think we should be looking for gradual improvement from about now - 7 January 2025.

And then I think we should hope - and demand from our government - more than the relief of simply moving to step 6 of the doom loop, only to start it again for 2025.

But what I also brought out in that piece was that in two areas of the UK, the state of emergency care was looking distinctly similar to 2022: and that it was Wales and the Midlands area that were showing particular strain.

I did it thirty-five minutes ago

It’s actually not until Thursday we get our next read on ambulance times for England (there’s an even longer delay for Wales) but from the handover and turnaround numbers that are trickling out, we can already get a good idea of what happened last month, and where we are right now.

Let’s not go on again about how the metrics of ambulance handover and turnaround times both directly predict ambulance response times, and provide a useful measure of overall NHS pressure. They do that, but going through it each time is exhausting. Let’s just look at the daily data (7-day smoothed to remove day-of-week effect) - and recognise that the recommendation is that all handovers are completed in 15 minutes or less.

Here’s the whole of England.

This is very bad in the sense that we are currently seeing handover times more than three times the recommendation, and the worst ever for this time of year (just).

And, to elaborate, we should expect these kind of handover times to result in mean response times for Category 2 calls (suspected heart attacks and strokes) of around 40-45 minutes in November for England as a whole. Again, this is about in line with the worst November ever, and are obviously massively unacceptable.

But we can remain optimistic that - at least at this national level - we are unlikely to head up that slope we saw in December 2022, when the Category 2 mean response times went over 90 minutes.

The reason for this is that we are seeing quite an early RSV and flu wave, which at least in the case of RSV may be peaking already. Early December indicators show a (little) bit of relenting in the pressure. And - more importantly - there is no sign of a COVID wave on top: so - with fingers crossed - we should be spared the very worst.

But now let’s just look at the Midlands.

This is qualitatively different. In the last few days of November, the Midlands have already seen conditions very much like the worst of the winter of 2022-3. And we have already had reports of one of the two ambulance services in that area changing their 999 call scripts to recommend that people try to make their own way to hospital - even Category 2 patients.

Pre-release data

So, some predictions. On Thursday 12 December, the ambulance response times for November 2024 will come out.1 The national headlines will look bad, but not disastrous - showing about the same times, or even a touch lower than the October read. Let’s say a Category 2 mean time of around 42 minutes. (Remember, the target is 18 minutes - throughout all this, we’re way out of the zone of what can be tolerated).

However, if anyone is curious enough to look deeper, I think they will find the real story - a disaster in the making (in fact, already made) in the Midlands. My suspicion is that - particularly in the West Midlands - we will see Category 2 mean times shooting up towards an hour, with the 90th percentile ones painting a more accurate, and even nastier picture: getting on for two hours. I’m not ghoulish enough to bet on this, but pretty confident.

There are - as always - some particular hospitals that will be struggling the worst here. I am not going to give details, since I worry about people reading this and attempting to drive while in no condition to, but we can make a few comments while keeping it high level.

1. It is the West Midlands that are suffering the worst. The East is approaching its worst state ever, but West Midlands is already there.

2. Each of the Royal Stoke, Royal Shrewsbury, Good Hope in Birmingham seem to be having a very, very tough time at present

3. Hospitals in the South West like Swindon’s Great Western are busy showing that just because the South West isn’t quite the worst area for ambulances any more - there is still plenty of room for individual trust crises.

Great Western’s landing page is almost entirely taken up with a massive banner that leaves no doubt about how difficult they are finding it at the moment:

Patronising advice

So, to re-iterate.

If you live in the Midlands, or in Wales, then for at least the next month, and more likely until March, you cannot rely on an emergency ambulance arriving in any kind of acceptable time. They will try their best, but on average, the services will be taking close to an hour to get to even those who are reporting intense chest-pain, or whose faces are drooping and cannot move half their body. There is a significant (>1 in 10 chance) that they will take several hours.2 The reason for this is that all the ambulances are parked outside hospitals, unable to transfer their patients.

There is no systemic alternative to an emergency ambulance. If you are vulnerable, or have any other reason to think that you might need emergency transport to hospital in the next month or two, it would be a very, very good idea to plan with neighbours, friends and family, in order to have an alternative method to get there.

In an emergency of course, dial 999 and follow their advice, but I would strongly advise asking the responder how long an ambulance is likely to be and whether they would recommend making your own way to hospital.

And yes, it is utterly insane to be repeatedly writing this in 2024, in a rich(ish) Western country, but - unfortunately - it is currently the simple truth about emergency care in the UK.

This Bigfoot report map, for example might be thought to give a possible location for the population of such creatures. And these might be thought to be confined to remote locations. However, the actual map does indeed seem to track the number of people who have the potential to invent sighting of Bigfoot, quite a bit more than it does plausible locations.

I was reminded of this group when looking at a report from the Preventable Deaths tracker - a project to bring together data hidden in Coroners’ reports across England and Wales, to make it analysable and draw insights from it.

At the moment, findings of avoidable or preventable death are openly reported, but only in unstructured text scattered through a bunch of verbose pdf reports, and trends are hard to find.1

The project aims to change this - a worthy goal.

Here’s a chart, from a recent post:

This is the number of cases in which the Coroner sent a report about a death with notes specifically addressed to an Ambulance service. This might give an indication of where the authorities - after appropriate investigation - have seen deaths that need investigation.

But - in the spirit of the Facebook Group - this doesn’t normalise for population (or, more relevantly here, the size of the ambulance service, in terms of the number of calls that each of these responded to).

And when you do that - take these coroner’s reports and simply normalise for how many calls are responded to by each ambulance service then the effect just simply goes aw…. oh.

Normalising for the size of the ambulance to does not make the effect go away. It does the opposite. Quite dramatically in fact.

Now, while this chart does align to the overall picture of the state of ambulance response in Wales being far worse than even the dire state of England, I do not believe that this is a simple measure of rate, or of the seriousness, of preventable deaths in the ambulance services. For a few reasons.

First, I may have made a mistake - the Welsh ambulance services reports everything differently to the English ones,2 and while I have done my best to get like-for-like statistics for “incidents responded to face to face” over the last 12 months,3 there is plenty of scope for error - I would appreciate informed corrections.

Second, we have a mismatch in time period. These reports are taken over a decade, and can take years after the death to come out. Most of the ambulance issues have emerged in the last few years (especially in 2021 and 2022, and these reports will take a while to dominate the numbers).

Third - and most importantly - I would very much expect these numbers to be swayed by the practices and activity of coroners in each of the areas. It is very plausible that in Wales there are simply more coroners who are more active/aggressive in asking for formal responses from Ambulance services when they see a case in which there are questions to answer.

Given the small numbers overall (only a few hundred cases over ten years!), this last is likely to be the dominant effect. And it may explain why the ambulance service with the best-looking ratio: in the South West, actually has the very worst performance in England in terms of ambulance response times.

In fact, these small numbers are perhaps the greatest surprise here. Looking at the times of ambulance responses - months upon months when even the median response time to a suspected stroke or heart attack could be measured in hours - it is abundantly clear that 46 coroners’ letters over ten years in Wales, and the even smaller numbers in England, does not even represent a scratch on the depth of the issue. They are exceptional tools in the coroners’ armoury, and are being deployed very sparingly indeed.

In the meantime the letters will continue to tick up. A few days ago, a new letter went out to the Welsh Ambulance service to respond to the case of a woman who waited 15 hours for an ambulance before her death - recorded of natural causes, but with questions to answer from the ambulance service.

The other obvious point to make is that even when these letters are written, they are almost all going to the wrong place. The ambulance services certainly have their issues, but at peak times, the main issue is that all their vehicles are queued up outside hospitals, waiting to unload their patients into overcrowded A&Es, which cannot safely care for them, because they cannot admit them into the main wards, which do not have a bed for them because they cannot safely discharge the patients they have. The response times of the ambulance services are just the outside indicator of fundamental capacity issues of the Welsh and English NHS.

These coroner's letters - an rare and exceptional mechanism that legally binds the recipient to respond with an explanation of the case, and what is to be done about it - should be going to the Welsh and UK Health Secretaries.

But for my money, the most quietly tragic life might have been lived by Hugh Everett III, the originator of the “Many Worlds” interpretation of quantum mechanics. This is the controversial interpretation, so beloved of science fiction writers, in which the whole world we see around us is recognised as the tiniest slice of a truly enormous objective reality - one in which every possibility since the start of the universe is realised. Everett came up with the overall idea as a young graduate student, and the detailed working through of the mathematics behind it became his PhD thesis.1

Everett’s tragedies were almost entirely self-inflicted. He was a coruscatingly brilliant student: in his first year at graduate school he published a paper on “Recursive Games” that John Nash2 considered extraordinary, and is now thought of as a classic of game theory. He studied quantum mechanics with Eugene Wigner, and wrote his PhD thesis under the guidance of John Wheeler.3

But even as a student, Everett was addicted to alcohol, food and tobacco, and he could not afford to indulge these whims on a junior academic salary. So even before his PhD thesis had been accepted, he took up a job in operations research at the Pentagon, ignoring Wheeler’s pleas to stay in physics, or at least come with him to Copenhagen and talk about his quantum mechanical ideas with Niels Bohr.4

Everett’s later work is almost all highly classified, but the fragments that have now been published give an idea of what he spent his life doing. These analyses are on kill ratios in the outcomes of nuclear war — how many would live and die in each stage of nuclear exchanges; whether the US or Russia would be able to absorb more casualties before completely collapsing as a military and political force.

We also know that Everett led the team that programmed the SIOP (Single Integrated Operating Plan) computers that would have launched the missiles to wipe out much of the population of Europe. They went operational in 1962.

Hugh Everett suffered from depression his whole life, and it’s difficult to think that his work helped. In any case, his depressive episodes got progressively worse, as did his addictions. He had married his wife Nancy in graduate school, but he continued womanising and seeing prostitutes, while growing obese and becoming more and more of an alcoholic. Nancy repeatedly cheated on him in return.

He was never very commercially successful. Some of the methods that he came up with: “attribute value” algorithms and the generalised Lagrange multiplier method, were later adopted widely, and made millionaires of some of his colleagues. But Everett himself was never able to fully capitalise on his insights, and his spending more than matched his income. Towards the end of his life, he was verging on bankruptcy.

But the greatest harm Everett caused was to his family. Somehow he never felt able to show affection to his children: Elizabeth and Mark, both of whom had their own struggles, and openly craved his love.

Elizabeth first attempted suicide in 1982. Mark found her unconscious on the bathroom floor, and got her to hospital just in time. When he returned home that night, he remembers his father looking up from his newspaper and saying mildly, “I didn’t know she was that sad”.

Everett himself died just a month later, at 51, of a heart attack brought on by his obesity and unhealthy lifestyle. Mark — then 19 — was the one who found his father’s body, and as he felt for a pulse, realised that it was the first time he could remember touching him.

In 1996, Elizabeth succeeded in her latest suicide attempt - her ninth - after being institutionalised for some years and undergoing extensive electroshock therapy. She left a note, saying she wanted to join her father in another universe.

Two years later, Nancy died of lung cancer, almost certainly due to second-hand smoke from her husband.

Mark Everett is still around. In fact, you have probably heard of him. He usually goes just by “E” and in 1995, he formed a band, where he serves as the front-man and driving force: The Eels.

In recent years, Mark has opened up more about his father, and has given historians of science access to his family’s papers. He even uses his real name from time to time.

(Image Source: Eels - official Band Website)

So, if you have ever wondered where song titles like “Elizabeth on the Bathroom Floor”, or indeed an album title like “Electro-Shock Blues” comes from … well, that’s where.

One of the things that turned up in Mark’s recent looks through his family boxes is a 1977 dictaphone recording of his father talking to Charles Misner, an old college friend and expert on General Relativity.5

This is eerie for historians of science to listen to, because in 1954, it was a drunken conversation between a much younger Everett and Misner that gave Everett the first ideas that he went on to develop into the Many Worlds interpretation.

In the recording, the two reminisce about the origins of these quantum mechanical ideas, while in the background, you can hear Mark practising on the drums.

1. Around spring, the incumbent government / devolved government promises that they are working hard to fix NHS capacity before next winter.

2. In summer, the relevant NHS bodies warn that nothing meaningful is happening on the ground and/or everything’s much worse, and they are in no way ready for the winter.

3. Through October and November we watch nervously as ripples of seasonal respiratory infections (e.g., RSV, flu and now COVID) grow into waves. The NHS know what’s coming and try to vaccinate as many of the old, young and vulnerable as possible.

4. As winter progresses, wards fill up with patients, so A&E is unable to admit patients to the hospital, so A&E fills up, so ambulances are unable to transfer patients safely to A&E, queue at the front door and so are unable to respond to new calls. Patients suffering from heart attacks and strokes wait hours for help, and incur enormous harms - enough to show up in excess death indicators.

5. In late December or early January, the waves of RSV, flu and COVID peak. If they all are high and peak at the same time, you get the winter of 2022 - when we saw a monthly average response time to stroke/heart attack symptoms of 1 hour 32 minutes. If, by luck, the peaks of infection are lower, or offset from one another, you get the winter of 2023 - average response time 45 minutes. The target response time is 18 minutes.

6. The waves recede in January and February, the NHS cares for the survivors and we bury the dead.

7. Go to 1.

   ---

Welcome to winter 2024/5

Right now, in 2024, we are currently somewhere in the transition between Steps 3 and 4. Ambulance response times are - once again - acting as an indicator of the overall pressure on hospitals: all the pressures pile up at the front door, and they are stuck parked there, unable to unload and cannot get back to pick up new patients. Adding more ambulances makes the queue a bit longer, but does not change much else.

“Sticking plaster” solutions such as changing protocols so that ambulances do not need to transfer patients to an A&E bed, but instead cram them into “corridor care” so they can go and pick up the lady with the heart attack from her kitchen floor … are in train. But the fundamental issues remain.

For 2024, any new investment does not have any time to make a difference, we are pretty much set in to suffer through steps 4-6. Really, the only two questions left now are “just how bad?” and “where will it be worse?”

First, how bad is it likely to be?

October ambulance times have just come out, and they suggest we are looking at a 2023-like trajectory, if we are lucky.

Or, more clearly to compare years like-for-like:

Not the worst, but fully in the post-pandemic trend of overwhelmed NHS capacity, not getting anywhere near the 18 minutes target. If we put this trajectory so far, alongside measures in the NHS England respiratory virus tracker, we wonder greatly whether flu’s prevalence at this time presages a higher peak than last year, and whether COVID could come around for “another go” before Christmas. If either or both of these things happen, things could get worse. If they do not, we might keep things down to a 2023-ish level.

So, while it’s not entirely clear, what the out-turn will be, what is clear that the system has not been fixed, and - yet again - it is up to external, uncontrollable factors on just how bad it will get.

Where will it hit worst?

As for “where”, then in the last few years, this has been blindingly obvious. It’s always been worst in two places: in the South West of England, and in Wales (particularly in Swansea Bay).

But this year, their unenviable crown is being challenged in the worst possible way. By the Midlands.

Looking, for example, at East Midlands ambulance service response times, and going month-by-month, we can see in this area 2024 is already starting to look more like the winter of 2022 than any other year. This is unique amongst the ambulance services, and even in absolute terms, the East Midlands response are currently tracking longer than the South West. This does not often happen.

Of course it’s down to the same things it’s always down to - the delays handing over patients at the hospital door. They are the worst in the Midlands than they have ever been at this time of year - something like this.

In particular, there appear to be problems mounting in a few trusts: Royal Stoke University, Kettering, Northampton General are each worth bringing out - but there are others too - all tracking with far worse times in the 2024 compared to previous years (and yes, even worse than in 2022).

Now, whether the Midlands will continue to see the worst pressure, and become a worse place to be seriously ill than even the historical “champion” the South West is rather a moot point.

The point is that people of the Midlands are tracking to see their worst winter yet, in terms of emergency response. If you’re vulnerable, consider whether you want to take up the free flu and COVID shots to protect yourself, and - though it’s completely unacceptable to need to think this in rich western country - you might consider planning ways to get to hospital in an emergency situation that does not depend on an ambulance arriving in a timely manner.

And … Wales

Unfortunately - again - I do not think that the Midlands will gain the dubious distinction of being the worst place to be seriously ill in the UK this winter.

I’m afraid that Wales — once more — is looking worse and trending towards a full-on nightmare for 2024. Times are tracking far worse than last winter, and the latest read is already ahead of October 2022.

[Update: 27/11/24 - added with October 2024 figures]

Now, beware of comparing these times directly to the English ones. They are medians,2 not means, and the categorisation of Red, Amber and Green calls (the Welsh system) is somewhat different to the Category 1-4 system that England uses. Nonetheless, the Welsh “Amber” also includes suspected heart attacks and strokes, and gives rise to stories like the death of Dave Strachan in March 2022.

The extraordinary point - not brought out enough - is that Dave Strachan’s story is not at all unusual. To their credit, the numbers reported by the Welsh ambulance services are extremely detailed, so we can estimate exactly how many people there are like Dave, who report urgent, dangerous symptoms, and are still left waiting over nine hours(!) for an ambulance to arrive.

The most recent month: September 2024, there were 13,963 calls which were, like Dave’s, were classified as Amber priority. That is, they were not immediately life threatening (not breathing, bleeding out), but still a proper full-on emergency: for example: symptoms consistent with a heart attack or stroke, where delay is likely to cause considerable harm.

The 95th percentile time is a decent measure of “how bad does it regularly get”. It is the lowest time in the worst 5% of delays. And for these “Amber” calls across Wales the 95th percentile is scarcely believable: 9 hours 4 minutes 48 seconds.3

Which means across Wales, roughly 700 people per month - 22 per day - are experiencing the kinds of delay that Dave Strachan did. Perhaps not all of them would have had symptoms quite as bad as Dave’s, but enough to be categorised as Amber.

Dave Strachan’s story is remarkable simply because it is so horrific and yet - still, apparently without comment or consequence for the Welsh - or for that matter the UK - administrations,4 is happening multiple times a day.

The Wall Street Journal talked to him (and yes, it was really him, they verified by asking him to make a tiny bet using one of the accounts he controlled), and he said the following:

• He reasoned that the 2020 overperformance of Trump would recur, because people were still reluctant to tell pollsters that they were voting for him.

• He reasoned that “neighbour” polls would give a better result. These are polls where rather than asking who you will vote for, you are asked who you think your neighbours, or social circle will vote for.

• He tested this with some private polling using that question, and there was a massive swing in favour of Trump. The results, he wrote, “were mind blowing to the favor of Trump!”

So, first things first - amazing, and all props to Theo for his analysis, his confidence, his willingness to think outside the box, and of course sticking through to a massive win.

But there’s a problem here. Theo is not the only, nor the first person to think about “neighbour” polls. These - and similar techniques for asking people about things which might be socially awkward to admit to1 have been around for decades.

They are not often used in run-of-the-mill political polling, because the answers are not usually terribly controversial. Simply: if a candidate has done well enough to be included in a pollsters’ screen, and it’s a one-on-one conversation, it’s not hard to admit.

But they are sometimes.

Radio Gaga

For example, here’s an NPR episode, from 4 November where they gave this method a shot with one of the best pollsters: Marist (they run the New York Times polls) and here’s the result:

At the time, the direct question “Who will you vote for?” the split between Republicans and Democrats was roughly even. […] On the friends-and-family question, the corresponding numbers were Republicans +10 and +14. [All voters, and “certain to vote” voters]

Report from Jay Livingston of Monclair State University

So, why didn’t everyone jump up when they saw this, and bet the same way as Theo?

Well … because this was for the mid-terms in 2022. (Yes, I was being mischievously misleading above - it was 4 November 2022).

And the results were completely different: there was absolutely no miss under-calling Republicans in 2022. It went the other way. The +10 and +14 for Republicans that the “neighbour” polling gave was far less accurate than traditional polling questions in predicting the true result.

If he had used his same logic in 2022, Theo would have bet massively … and lost it all.

What about 2020?

Another team did tried the same technique back in 2020, and got some better results.

Read more

Still, the dynamics provide a clue as to the underlying trends, and with something like whooping cough, where the “whoop” is so distinctive that experienced GPs can diagnose it over the phone, it’s a particularly good guide to how many cases there may be around, even if a few may turn out to be false alarms.

These reports have been offline for a little while, as the UKHSA experiment with new ways of collating and publishing the results, so we haven’t actually seen anything since the summer.

But now, while the weekly tracking is still suspended, we do totals up to week 44 (i.e., up until last week, end of October).

For things like whooping cough. And … well … at first sight at least … wow:

Astonishingly, and despite what the chart looks like, this is fairly good news.

The reason this is good news is we already knew that this year had seen a quite absurd spike in whooping cough in May. Looked like this.1

And now we have updates for the totals of the 44 weeks to the end of October 2024, so we can take the difference, and estimate what the interim values would have had to be. The line won’t be quite right, but the average is.

Something like this:

So, this is only fairly good news. We are now ongoing with whooping cough cases at around 10x the rates they were in mid-2023, but at least we do not have them at 100x, which is where we spiked in mid 2024.

The trouble is this is an unstable situation. The achingly low rates of vaccination we are see (particularly in London and parts of the Midlands) implies that we should expect this new higher level of cases to continue, and new spikes to jab up every now and then, and harm young kids on a very regular basis.

Likewise, measles. We have come down from the spikes of late 2023-4, but we are still seeing higher levels ongoing, and will have new spikes with seasonal rhythms, as well as local outbreaks, each one of which has the potential to spread further.2

The worry is that it will take a truly huge outbreak of either measles of whooping cough - of the scale that we have not seen for decades - to do something astonishingly brutal. That is, for an outbreak of some vaccine-preventable disease to kill a sufficient number of children that people finally listen to public health messages and get themselves and their kids vaccinated at the 95%+ rates that you need for population-level suppression.

It would be really good indeed if we could come up with just about any way of getting to the needed vaccination rates that does not involve this “too many dead kids to ignore” step, but - so far - nothing seems to be working.

They weren’t. It’s true to say that they had errors pretty much in line with historical levels1 … but that is to say: they were really awful. Here’s the top-line/popular vote miss.

And then you drill get down to the numbers that actually decide the result: the individual state-level vote shares. The errors here was were also large, and worse - they were highly correlated in one direction.

To illustrate this - let’s see how the final polling margin averages2 compared to the result (so far) in the individual states. Here’s a visual look at them (Showing Trump minus Harris margin in each case: polls = start of each arrow, result = end of each arrow). Red where the polls underestimated Trump, Blue where the polls underestimated Harris.

The trouble with this is that the errors don’t look too enormous in comparison with the differences between the individual states. So let’s take an objective measure.

Error metrics are a big topic in statistics and its various aliases (machine learning, deep learning, AI), but let’s go for the one this Substack is named after: the “Mean Squared Error”. That is, the average of the differences between the estimate and the result, squared and then averaged. We often take the square root of this number, so that the units are the same as the thing you are measuring - in this case percentage points.

So, the (square root of) the mean squared error of the 563 state-level results here is 3.7ppts. Which is a reasonable measure to give as the “average” state level polling miss.4

In terms of being any decent guide to the overall result, you have to say that this is at least unsatisfactory. The seven “swing” states - the ones that were always going to decide the election - were all polling within one or two percentage points of going one way or the other.

So a miss with a (root) MSE of 3.7ppts, very highly correlated in one direction, this simply means that all the swing states would go that way. Which they did - to Trump.

No, I mean BAD

But - the polling defenders might say - this is simply an unrealistic standard. It’s not the error itself that’s the issue here, it’s how close the race was. You just can’t expect any methodology to do much better than 3.7ppts RMSE, and we’re just unlucky this cycle that the race was so close that this kind of miss was critical.

OK. Let’s take that Mean Squared Error measure, and instead of comparing the final 2024 polls to what they were supposed to be to giving a guide to: the 2024 result, what happens if you compare them to the 2020 result instead?

Now, let’s be clear - none of the polls were meant to be measuring 2020 results, or even be a good guide to it. If anything they are meant to tell us how much things had changed since 2020. It just gives a convenient benchmark for how much they’re managing to give an accurate view of this election versus any other one. The

Aaaand, the root mean squared error of the state level 2024 final polls in predicting the 2020 results is better. Much better: 2.4ppts.

This is a little absurd, so let me emphasise the point: the final polls for 2024 - as aggregated by competent modellers - were a better guide to the 2020 Presidential results than they were to the 2024 results.

That’s not fine.

Rather, it gives a hint that in that storm of weighting, sampling and other odd-looking operations pollsters have to get up to in order to produce a balanced sample … all appear to work better to reproduce the last result than they do to give an accurate read on the current state of play.

People often criticise political campaigns for their mistakes in “fighting the last war”. In this case, there is a clear explanation for this: the polling that is supposed to be giving them a view of this years war is better - literally - at showing them one from four years ago.

There has been much written on some of the individual markets, and even individual traders, who might have been behind this move. We could not really distinguish between four hypotheses about what was going on.

1. Genuine belief that Trump is under-priced from public information (i.e., that the true probability of their winning is significantly higher than 50-ish% percent).

2. Genuine belief that Trump is under-priced from private information (e.g., unrevealed scandals or candidate health issues that sway the result).

3. Market manipulation for political reasons - believing that it will help Trump to be seen as further ahead just a couple of weeks out from polling day.

4. Market manipulation for financial reasons - a series of so-called “pump and dump” moves

Add two days to the time-period above - let’s look at “Kalshi”: the largest betting market that’s legally open to US residents

Down from a peak of 64.6% to 55% today. A ten-point move.

The other markets are making similar moves. Polymarket, which was up at 67% for Trump on October 30 (and subject to the huge “Fredi9999” bets in his favour), was down at 57% this morning.

That is, in the last few days before the election, the betting markets are swooping back from their mid-or-high sixties towards the 50% line: which is where most of the polling and the betting models are waiting for them.

Why are they plunging … Pumping and Dumping?

Obviously, this could be new information. Several of the models have ticked a few points towards in favour of Harris in the last few days - notably The Economist’s model now (November 2nd) has her narrowly in the lead for the first time in two weeks. But no-one has a swing of 10 percentage points of probability in the model. Nothing either close.

It also could be more market manipulation: either Democratic operatives concerned that Harris was “looking like she was going to lose”, or from Trump supporters worried that if Trump looked like he was sure to win, then his supporters might not want to vote. But we can put absolutely anything down to market manipulation - and this time around there’s simply no evidence for it.

Read more

In England and Wales, the authorities do not even allow a death where the coroner is investigating a cause to be entered in the statistics, so you can get very strange patterns. This is a particular problem for deaths in young people, because many more of these deaths need investigation, and so those statistics sometimes do not settle down for months or years.

What this means is that if we are looking for medium-to-long-term effects from the pandemic, then even now, more than four years out, UK statistics have to be taken with care.

Scotland to the rescue

In Scotland, process is much faster; they register deaths straight away, revising later if need be, so we can look to their statistics (at least to end of 2023, when all revisions should be complete) to answer both reasonable, and less reasonable questions, such as:

1. Has the much-feared effect of lockdowns from 2020, causing people to miss cancer diagnosis and treatment, come to pass?

2. Was there anything in all the noise from the anti-vaxxers? E.g., the “turbo-cancer” claim that there has been a shocking rise in cancer deaths since 2021, when the vaccinations started.

Here we are overall - cumulative (to all causes), and then non-cumulative, to bring out the major components:

First impressions.

1. It’s very seasonal (except cancer). Respiratory, circulatory, even dementia. Scottish winters are hard.

2. It’s going down over the long-term, but overall improvement steadied and slowed to a halt from around 2014 and this is driven by a rise in “other causes”

3. COVID death rates were shocking, but thankfully brief (and note this is only counting where COVID was listed as an “underlying cause” of death, not where it was just mentioned on the certificate).

We can get rid of point one, and bring out points two and three are more strongly, if we smooth over 12 months to show the trends (though here we are also “smoothing out” the COVID spikes over a year)

We see that “Other cause” turn around in 2014, and rise. It is behind the overall lack of improvement. Dementia and Alzheimer’s disease has also been rising across this period, though this has steadied and even dropped recently.2

“Other cause” in Scotland

This “Other cause” category is well-studied in Scotland. The rise in death rates from 2014 was - to a large extent - was driven by drug abuse, specifically opioids. Only Scotland, the US and Canada saw such a rocket in 2014 in drug-related deaths: a full-on opioid epidemic. A chart from Colin Angus at the Sheffield Centre for Health and Related Research (SCHARR):

There is a lot more to say here about “a lost generation” (Scots - especially men - currently in their mid-40s) and on cultural and deprivation-based differences. Some of it has been said, and not much has been addressed. Although the generations are similar, it’s notable that the causes do not seem to be the same as they were in the US, with supplies being less through the abuse of legal routes by pharmaceutical companies, and more full-on illegality (opioids abused in conjunction with street benzodiazepenes).

What about Cancer?

One line that doesn’t seem to be doing much is that quiet mauve one, starting at the top and then reducing over time. It certainly looks as though it’s just might be declining overall.

But we can look closer. By age3 and sex.

Again, all kinds of interesting health and social lessons to be drawn from this (older Scottish men urgently need to stop smoking and eat better), but the overall trend can be seen to be in the right direction, particularly in the younger age groups.

The trends though, are clear - at least in the case of Scotland, any of those feared effects of delayed cancer diagnosis and treatment do not seem significant in the overall mortality numbers,4 or at least so very delayed that they have not shown up in deaths three years later.

And as for any epidemic of “turbo-cancer” starting in 2021, it is unsurprising to see that something invented by self-promoting charlatan quacks to sell vitamin supplements does not show up in the real world, but it’s reassuring to see nonetheless.

The state of the nation

As I’ve previewed a few times, I think the commentariat is in danger of confusing an extremely uncertain race with an extremely close race.

Some things I think are true:

• I think that most of the pollsters are doing their best to get things right, with solid methodology; and so the polling averages are as good as they can be. Rasmussen, AtlasIntel, Trafalgar and perhaps some others aren’t doing their best to get things right - they’re doing another job. But adjusting for this only shifts the average a tiny way towards Harris.

• I think pollsters are making an error in starting to weight by recalled vote - an over-reaction to the failures of 2020 and 2022. It might help, it might not, but I think it’s just adding noise and uncertainty at this point. If this is right, it results in a slightly larger tilt towards Harris, but we’re still in toss-up territory here.

• I think the methodologies will turn out to be systemically wrong in one way or the other to historical levels (around 2ppts or beyond, and more at the state level), and because these errors will tend to turn out to be correlated across states, one side or the other will win by more than 100 EV. I put a very low probability on the errors being uncorrelated.

As a result, my expected outcome looks something like this:

Translating into an electoral college expectation

So, what does this mean when we translate to Electoral votes?

Most people would say that we only really need to look at the swing, or “toss-up” states. That Democrats have 226 EV banked, and Republicans 219, and it’s only the remaining 93 that are up for grabs. I don’t quite believe this, but I am willing to accept that the swings that you’d need to get states like Minnesota, Texas, Virginia and Florida to flip outside their historical voting patterns are in the 20%-or-below bracket.

I’d also note that a full-on blow-out landslide scenario is more likely for Harris than Trump, just because both Florida and Texas (70 EV total) fall into the range of a large (5-7ppt) polling error, while Harris has about 31 EV at the same level of risk.

So, back to the swing states: there are seven of these, worth 93 Electoral Votes.

Here are these seven states, and the latest polling averages (via 538 - others give similar values).1

So, given that a reasonably correlated 2ppts swing in either direction would allow Trump or Harris to sweep the lot, that’s what we should expect. That’s just the kind of error we tend to get in US Presidential elections.2 There’s only one possibility in the last six cycles that would even give us a close race, and even then, only if the error broke towards Harris.

So I think we’re about at:

75% - Convincing victory - winner gets 300 EV or more

25% - A close race - winner gets 299 EV or fewer

Of for ****’s sake, who’s going to win?

But although all this about the closeness of the race is helpful, and perhaps even relevant to important events, given the potential for close votes to be contested, it’s not what most people want to know.

What most people want to know is who is more likely to win: Harris or Trump, and hang whether it’s in a convincing victory or not.

Up until last week, I would have said (just) Harris - I think the questionable pollster methodology is more likely to be slanted away from her than it is away from Trump. And so a significant error is slightly more likely to lead the actual vote to break towards her than to Trump.

But now early voting numbers being reported by Jon Ralston over in Nevada (the only early voting commentary anyone should pay any attention to3) have shaken this belief. And shaken it hard. If we take his reporting at face value - and I think we should - then Harris is struggling in Nevada and falling well behind in the early voting. That is, so far she is doing significantly worse than the polling in that state suggested. Now, Nevada itself is not a huge issue - it only represents 6 EV. But if we apply the idea that polling errors are likely to be correlated across states, then the consequences are fairly obvious - she is much more likely to be struggling behind her polling in other states too. That’s how conditional probabilities work. And very quickly, Harris looks to be in trouble.

On the other hand, the early voting in Nevada is only a weak signal - it may well just be showing us a realignment post pandemic of how different types of people vote. We will know a little more soon, and a good get-out-the-vote could swing it back on election day, but I think we must take it seriously enough that these numbers significantly shift the overall chances.

Given all this, I think now the chances now look something like this:

• 50% on a convincing victory for Trump

• 25% on a convincing victory for Harris

• 25% on close race, in which polling errors aren’t well-correlated - and we can cut this down the middle.

Overall, this gives a 62%/38% split in favour of Trump. Which is - coincidentally - not far off where the betting markets are.

Now, these are not the probabilities I’d personally want at this stage, and to be clear, they’ve switched considerably in the last few days (since seeing the emerging early voting patterns, in fact).

I also don’t claim to understand anything about how American voters think. In fact, these patterns make me even more sure of my inability to comprehend than I was before. I just think that this picture is close to where you end up objectively, if you take seriously the historical patterns, the most likely dynamics of the polling errors, and what we know about the early voting patterns.

UPDATE: 4 November: Jon Ralston in Nevada has gone full reverse ferret. Despite huge differences in favour of Republicans in early voting patterns, he is clear on his final prediction.

So, what does this change? Well, Ralston’s prediction - if we take it as gospel - would represent a 0.5ppt-1.2ppt swing towards Harris versus the polls (depending on who’s polling average you take).

And if I continue to rigorously apply the logic I’ve been applying so far (there will be polling errors, they will be significant, and they will tend swing in the same direction), then this simply reverses that tiny push at the end taking me from one “hump” of the probability distribution to the other. An error of that size is slightly smaller than I’ve been expecting, but if it breaks towards Harris and is correlated across states would be enough for her to get over the line in a fairly convincing manner (especially as the polling averages have also slightly shifted towards her in the last few days).

So, do I want to take this tiny sign from one analyst - Ralston - in a state with a grand total of six Electoral Votes, and have it reverse everything in the last few hours before polls close? Or do I want to stick to my guns and stay where I am?

It’s pretty much a choice between being consistent on methodology, and consistent on answer.

I’ll be consistent on methodology. I still think the overall distribution is bimodal, I still think the errors will be correlated. And I am still extraordinarily swayable by the tiniest signs that the break is going one way or the other, which - honestly - will make these last hours one hell of a rollercoaster. The terrors of a bimodal probability distribution.

So, Uno-Reverso we go:

• 50% on a convincing victory for Harris

• 25% on a convincing victory for Trump

• 25% on close race, in which polling errors aren’t well-correlated - and we can cut this down the middle

And don’t believe anyone who says they are sure.

1. How far the polling margin average would have to move for the race to flip.

2. How far the polling margin average typically moves in the last 12 days.

3. How far the polling margin average typically is from the result.

Get these three clear, and you can start to see what the next 12 days are likely to look like.2

The polling margin average would need to move around 0.6 points to flip the result

Right now, the headline polling averages are that Harris is ahead of Trump by about 1.6ppts in the headline popular vote.

(Silver Bulletin averages)

The issues is - this isn’t enough of a lead for Harris. Not nearly

The Dems have what’s called an “Electoral College" disadvantage”. Models differ of the size of this, but most would align that - if you make certain assumptions about swing states moving together, Harris needs to be ~2.2 ppts ahead - in absolute margin - in the national popular vote to expect to get more than 270 Electoral Votes. That is, she needs an absolute vote margin swing of around 0.6ppts in her favour to take it in her favour. Or - to put it another way - Trump is about 0.6ppts “ahead” in the overall race, due to the way the Electoral College is counted.

The polling margin average tends to move move by about 3 points of margin in the last 12 days of a Presidential Race

Right, if that’s the “lead” then take this chart, or something similar to it, and look 12 days out. The typical movement in margin in the polls is around 3ppts in the last 12 days.

(FivethirtyEight)

So, the typical (and so, arguably expected) movement in the next twelve days is approximately four times the margin that Trump is currently ahead in the Electoral College.

And this is just the polling movement.

Read more

Most of the models have the race as a dead heat, or within a percentage point or so of the mid-line. And over history, the betting markets have largely been following their consensus. That is, if you shaded the area between the highest and lowest estimate from the models for each candidate, you’d find the betting markets in this range. Which would make sense if both they, and the betting odds, were reacting to the best information available.1

But the betting markets, starting on about the 7th October, diverged from this range covered by the models. This move was led by Polymarket - where close watchers of the betting quickly realised that one large account was driving the majority of the betting for Trump. This pushed his odds up to around 62%, before falling back to the high fifties.

Other markets followed a little later, and in less extreme ways. The series of moves could simply be other traders reacting to the arbitrage opportunities. But the original market trader remained in the background and added further to their positions, generating a great deal of speculation on their motivation.

Why are they doing this?

This saga has been well-documented by Rajiv Sethi in a set of articles that run through the evidence for an original “Fredi9999” account having expanded to several others, with at least $25m on Trump. There is also (rather circumstantial) evidence of some of these accounts being linked to a person going by the name of “Michel” who appears to be a French national, claims no political allegiance either way, but claims that it is clear to them that Trump’s probability is closer to 75% than to the prevailing market price.

Imperfect Information

The French Connection

Over the past week or so, a sizable gap has opened up between prediction markets and statistical models referencing the outcome of the November election…

Read more

2 years ago · 19 likes · 4 comments · Rajiv Sethi

There seem four possible motivations:

1. Genuine belief that Trump is under-priced from public information (i.e., that the true probability of their winning is significantly higher than 50-ish% percent). This is what “Michel” claims, denying any political motivation or special access to information.

2. Genuine belief that Trump is under-priced from private information (e.g., knowing about internal polling, unrevealed stories, strategies, candidate health issues that are likely to become public in the next few weeks, that would tilt the race towards Trump).

3. Market manipulation for political reasons - believing that it will help Trump to be seen as further ahead just a couple of weeks out from polling day. And so putting money into the market to spread that narrative.

4. Market manipulation for financial reasons - the so-called “pump and dump” move, where you put in enough bets that the “trend becomes the story” - others pile in assuming you know more than them (i.e., they assume that you are operating for reason '#2) and then you sell at a profit.

When major outlets have written this up, they tend to emphasise one motivation or the other.

The Wall Street Journal - putting the number above $30m - lean towards #3 - market manipulation to make Trump appear further ahead than he actually is for political reasons (though what these are isn’t entirely clear).

The Financial Times writes up the same situation, but leans towards #1 - a genuine belief that Trump is significantly underpriced. Sethi himself seems agnostic, but leans somewhat towards #4 - the possibility of a “pump and dump” motivation, though Fredi9999-and-friends seem to be holding on to their positions for now.

Also, recall that something like also happened in 2012, when the so-called “Romney Whale” was responsible for around a third of all trades made in favour of Mitt Romney to win, and kept the contracts right though to election day (they are estimated to have lost around $4m in doing so). Professor Sethi was the one who most comprehensively analysed this single trader in a series of academic papers (which also touched on various theories of market structure) - hence his interest in a 2025 re-run.

So, what is going on?

Well, clearly and most obviously, we can be fairly sure that there is currently a premium on Trump on several betting markets driven - most probably - by the activity of one single very deep-pocketed agent.

At a reasonable guess based on the order book that is made public, this premium is around +5ppts more on Trump versus where the market would settle if this one agent’s trades were removed. And this would take the price back into the low 50s - into the range shown by the models.

There is no evidence that Fredi9999 has privileged information (e.g., internal information from either campaign), though right now they are doing well in terms of being ahead of the crowd. Since October 7th, the polling has moved a few fractions of a point towards Trump, so if Fredi9999-and-friends liquidated their positions right now, they would make a sizable profit - hundreds of thousands of dollars. However, given the size of their bets, it’s likely they would not consider such a profit a substantial one.

So, given this evidence, I would suggest that betting markets are distorted by one particular participant, who has no particular reason vs any other person to be right about the chances in the upcoming election.

But they do have a lot of money,2 a set of beliefs about the result that is out of the usual range, and they want to bet on these beliefs.

The motive (almost) doesn’t matter

If all we are interested in is the clearest, most faithful reflection of the truth about the state of the race, we don’t actually need to determine which reason is correct. Or, to be more precise, we don’t need to distinguish between #1, #3 or #4. In all these cases, we should just discount the betting markets by (most of) 5ppts and continue as normal.

Just to quickly see why: considering reasons #3 and #4, we have a deliberate attempt to distort the markets as an objective aggregator of information, and this attempt is succeeding. We should discount it.

In the case of #1, we do not have deliberate attempt to distort anything - it is a genuine belief that should be counted in the “wisdom of crowds” interpretation of the betting markets. The issue is, it is being counted with far too much weight, because they are putting $25m-$30m into the markets based on it. Whatever your belief about the weight we should give to richer people’s opinions, it should not be literally tens of thousands of times more weight than others. This is particularly true when - most likely - the source of their wealth has little-to-nothing to do with predicting political races.

Therefore, we might want to weight this belief a little more than other participants (because of their willingness to put much more money against them), but nowhere near how they are being counted. And the first approximation to this number is - again - to remove their distorting presence frm the market. Which takes us back to taking about 5ppts off the price while these positions continue to be built.

The only case in which we should not discount the “Fredi premium” is #2. If we believe that this person has privileged information that justifies the amount of money they are willing to put at risk. And - frankly - the evidence for this is thin-to-non-existent. Internal party polling tends to paint a very similar picture to public polling,3 so having access to this information wouldn’t change anything much. A hidden scandal, or concealed medical condition of one of the clients would be a different matter - especially if there were some reason to think that this was going to be made public before November 5th, but one would expect this to have come out by now for maximum damage. There is a reason it’s called an “October Surprise”.

So, tentatively, I think those engaged in truth-seeking have a clear way forward. We discount all, or almost all of the Fredi-premium (say - 4ppts out of the roughly 5ppts total4), and view the betting markets as fundamentally communicating a ~53% chance for Trump / 53% for Harris, which puts them back into the range of the indicators we spend most of our time looking at - the poll-based models, which are showing almost exactly these numbers at the moment.

There are all sorts of suggestions out there: differential turnout, “the vibes”, early-voting patterns, attendance at rallies, leaving early at rallies, voter registration information - all being held out to make it look absolutely certain that all the uncertainty in the polling is an illusion, and that either Trump or Harris is going to romp home in a landslide.

The thing is, most of these issues are made every cycle. It doesn’t take long to find the screaming headlines and breathless predictions from 2008, 2012, 2016 and 2020. And after the election they are never mentioned again, because never seem to be anywhere near predictive of the actual results as - you know - the polls.

Amist this hurricane of blowhard rhetoric, there are four factors being discussed which have at least a hint of a point this time around. These are four that I think deserve a bit more consideration than a simple eye-roll, and a look at the numbers you’d get if you accounted or adjusted for them.

In order of plausible impact on the race (my judgement), they are:

1. The track record (pollsters always underestimate Trump, don’t they?)

2. Flooding the zone with shit (the partisan pollsters spewing out unreliable numbers, aren’t they?)

3. The ground game (surely more money, and more physical presence will have an effect on turnout?)

4. Total Recall(ed vote) (a controversial polling methodology change - is it right?)

Let’s look at these, in ascending order of plausible impact on our picture of the race.

The track record

Bluntly: in every Presidential election Donald Trump has fought, the pollsters have under-called his votes, and done so by historically large amounts. Look at 2016 and 2020.

However, the inference: to think that this means the polls will undercall him in 2024; is undermined by this very same history. Pollsters try to get it right,1 so when they’ve missed they tend to learn from their mistakes for next time. They compare their results to the actual results, diagnose the problem and adjust their methodology. You can see this effect in the chart. There is no particular pattern of systemic undercalling that lasts from one year to another. They course-correct and the error next time is not predictable from the error last time.

If anything, when pollsters miss one year, one should be at least as worried about an “overcompensation” next time around, as you are about a repeat of the miss. The chance of a polling miss is large. But its direction is as unpredictable as ever.

Does it make much of a difference? Yes. Misses of the order of 3.0 or 4.7ppts would absolutely transform the race from a toss-up, to a nearly-landslide victory one way or the others.

Is there a case for an adjustment? No. The same history that tells us that there is an effect, also tells us we can’t tell which way to adjust in 2024.2

Net effect - zero.

Flooding the zone with shit

In 2022, one of the reasons there was a narrative for “a red wave” - i.e., a Republican sweep, that never actually materialised, was a slew of polls from heavily partisan pollsters which overestimated Republican share, especially in Senate races. And by large margins in some cases: 5+ points ahead of where they actually ended up. Whether this was wholly deliberate or not (projecting that your candidate is way ahead is - at best - a double-edged sword), it had an effect.

And, say many, the same is happening this year, with pollsters like Rasmussen (who are demonstrably working to try to get Trump elected) releasing poll after poll showing Republicans ahead. This is just that wave of shit flooding the zone again. Ignore it.

Should we take this seriously, and will it have an effect?

It seems clear we should take this seriously. If you have pollsters who care more about a political outcome than getting the result out, they will add to error. This is a real thing, and should not be dismissed.

The thing is, most of the poll aggregators already exclude (or adjust) the more egregious of these partisan pollsters. Rasmussen and Trafalgar are already banned from most models, or at least heavily adjusted for past bias before being included.

And the rest … well, the issue here, is that this time around they’re not actually saying anything very different to the better-quality pollsters. Looking at the national vote, and some of the key states, I did a rough piece of work to rip out the more obviously partisan guys (which agrees, roughly, with a similar analysis done by Nate Cohen of the New York Times). Here are my results - his at the link.

So … honestly, this is hardly any difference. Not because there’s not a lot of Republican-leaning pollsters (there are, alongside rather fewer Democrat-leaning ones), but because their current results are very close to the best-quality ones.

They may be flooding the zone, but they’re not flooding it with anything very different to anyone else.

Does it make much of a difference? Not at the moment. Maybe two or three points of probability to win the Electoral College (because of Pennsylvania, mostly).

Is there a case for an adjustment? Yes. But most of the polling aggregators are doing it already, by excluding or adjusting the worst offenders.

Net effect - minimal-to-none.

The ground game/get out the vote

There is much buzz in the press about the apparent energy and money going into the Democrats’ field operations, and the lack of the same into the Republicans’. This has been written up in many places, both with obvious partisan lean and more mainstream news operations.

If the indicators here are true, then Democrats are using a lot of money on a tried and tested grassroots-led local approach with field offices and plugged-in volunteers, and the Republicans are effectively outsourcing the whole thing to a bunch of loosely affiliated organisations who haven’t any track record, and mostly don’t like one another. If this all holds as is strongly implied in these articles, then come election day, this will translate into Democrats turning out in much higher numbers than the Republicans.

Should we take this seriously, and will it have an effect?

So, research on whether this will have an effect is somewhat mixed, with randomised trials on phone-operations coming out often as effective dollar-for-dollar as the more visible door-knocking approaches that Harris’ campaign has seemingly gone for. But if we believe that Harris’ campaign can move turnout amongst Democrats by just a tiny amount, it can have a big effect. Some point to the 2020 race, where Democrats - worried about the effect of COVID if they went door-to-door, cancelled the vast majority of their physical ground-game. Republicans did nothing of the sort, went ahead with their door-knocking, and overperformed the polls by an average of a whopping 4.7ppts.

+4.7ppts vs the polls would, of course mean Harris’ chances would jump from their current 50%-ish place to a touch over 80%.

But there’s reason to be (very) sceptical about this. First, while I believe the effect would be fairly significant, it’s not going to be the full 4.7ppts. Putting all of the 2020 polling miss down to ground effects on the day is … brave. But - one might argue - even half of this would be a big difference. And this is true.

Read more