They've done a lot of work on at least two dimensions: (1) handling the nonstandard sounds and habits of speech of very early readers, who might be as young as four, and (2) connecting this to a specialized teaching system based on the science of reading, e.g. decodable readers.

(I'm not sure whether Cato is generally reliable, but FYI there are lots of other writeups online on the same topic. It was the first non-paywalled and reasonably complete one I found)

[1] https://www.cato.org/blog/phonics-failure-public-schools

At the time of writing, the sentiment in this post is that this is a terrible idea, and that kids need human tutors. The latter is 100% true. But also, you might want to know some facts about the state of children's literacy in the US (Ello is a math and reading tutor):

1. We're in crisis. As of 2025, 40% of fourth graders are reading below basic levels [1].

2. There's a massive teacher shortage. 2025 US state data shows ~400k teacher positions either unfilled or underqualified [2] – over 10% of the workforce.

3. Bloom's 2-sigma shows that 1-1 tutoring delivers outcomes at the 90th percentile of classroom teaching. Early research is finding that AI can deliver some of this benefit [3].

4. This can't always be solved by parents: 54% of US adults have a literacy below a 6th-grade level, and 20% are below 5th-grade level [4].

At Ello, I heard stories of children figuring out they were behind at school, and when given the app, they holed themselves up in their room and used it to get themselves caught up. And then they could read! Can you imagine falling behind at this critical juncture, and being stuck illiterate while your friends grow past you? We're currently setting kids up for lives of shame and deprivation.

My take: this really is a life-changing technology. And we need this problem solved. Democracy doesn't function without an educated populace.

[1] https://www.nagb.gov/news-and-events/news-releases/2025/nati...

[2] https://learningpolicyinstitute.org/product/overview-teacher...

[3] https://www.sciencedirect.com/science/article/pii/S2666920X2...

[4] https://www.thenationalliteracyinstitute.com/2024-2025-liter...

Check the archive.org link at the bottom!

We didn't get around to addressing this within the paper itself – 80 pages is a lot, and deadlines, etc. But I have unpublished experiments that show that in a reasonably broad setting I'm doing some work in, verbalized probabilities are restoring a distribution that looks almost identical to the base distribution. It is not possible to demonstrate this on frontier models, since their public models are already mode-collapsed, and they don't share the base model or logprobs anyway. But I've established this to my personal satisfaction on large local models which offer base / post-trained pairs.

To share some intuition on why one might believe this is occurring: there are a bunch of tasks implicit in the pre-training corpus that encourage the model to learn this capability. Consider sentences in news and research articles like: "Scientists discover that [doing something] increases [some outcome] on [some population] by X%". It seems quite natural that the model might learn a pathway by which it can translate its base probabilities into the equivalent numeric tokens in order to "beat" the task of reducing loss on the "X%" prediction. I can even almost visualize how this works mechanically in terms of what the upper layers of an MLP would do to learn this, i.e. translating from weights into specific token slots. And this is almost certainly more parameter-efficient than constructing an entire separate emulated reality for filling in X. Although I'm not ruling out that the latter might still be happening – perhaps some future interp research might be able to validate this!

I'm actually working on a paper that packs up some of the above findings in passing. But if helpful in the meantime, this is also building on related work by Tian et al. 2023, "Just Ask for Calibration" [1] and Meister et al. 2024, "Benchmarking Distributional Alignment of LLMs" [2], that give some extra confidence here. Their findings indicate that whether or not they are rooted in the model's base probabilities, they seem to be useful for the purposes that people care about. (Oh, and you can probably set up an experiment to verify this independently with vLLM in a few Claude Code requests!)

Hope that was helpful – feel free to ping with follow-ups! (Although replies might be a little delayed, I happened to see this at a good time; having quite a crunchy week)

[1] https://arxiv.org/abs/2305.14975

[2] https://arxiv.org/abs/2411.05403

Sadly, within the taxonomy itself the authors restrain themselves from sharing their findings on the most delicious parts and preparations of the animal. Darwin would have been disappointed [1], but at least as a species we've gotten our time down from 300 years [2].

1. https://www.npr.org/sections/thesalt/2015/08/12/430075644/di...

2. https://www.youtube.com/watch?v=zPggB4MfPnk

This is such a new capability that I'm having a hard time getting a sense of interesting use cases. I'm quite sure this is more than just a shim for web services which don't expose APIs. But I also wonder whether LLMs are good enough to be trusted with more open-ended tasks at this stage.

What a shame it would be for this culture to be lost; while there's a lot of dross in the blogosphere, I don't know if the brightest jewels will still be possible in a future system of local, private, transient clusters of thought.

> CDC recommends collecting three stool samples from patients over several days for accurate test results. Commercial testing products for diagnosing giardiasis are available in the United States. [1]

Perhaps running three tests is the standard of care, or if not one might advocate for this based on the CDC recommendation. And if dismissed, perhaps there are commercial products available at the consumer level.

[1] https://www.cdc.gov/giardia/hcp/diagnosis-testing/index.html

In my experience the medical system is unusually useless and dismissive with digestive issues. I think this is probably related to how little it can do in this area. 10-15% of the US has IBS, and this is a disease of exclusion. That literally means that the medical system acknowledges a cluster of symptoms, but has no idea what is causing them or how to cure them. I can imagine that blaming patients is easier than the alternatives for some doctors.

I was fortunate that over time I managed to return myself to full capacity, through reading a ton of research and running dozens of experiments like the above. But it was so damn hard. The symptoms reduced my ability to use my brain to fix myself. And if you're not a careful eater, it's not at all intuitive which foods contain gluten. This was also almost a decade ago while living in a developing country, so it wasn't even apparent that gluten might be a suspect.

I'm currently based in the US - does anyone know how one might get properly tested for chronic giardiasis, as a person who isn't themselves in microbiology? I almost certainly encountered poorly treated water in that period of my life.

Also - I can't help but suspect that a nontrivial percentage of the developing world is living below their full capacity due to something like this. Neglected tropical diseases are a horrendous category.

I'm currently feeling out a little research project around the use of AI in improving group decision-making, specifically with the hope of improving our political systems. There are so few IRL case studies, it'd be great to have more intuition for this!

(I'm @dch on twitter if you'd prefer to DM)

I wonder if that suggests this is one of those universal pre-linguistic words, like "huh" [1]. That feels especially believable given the utter simplicity of the mouth sound. It's almost just a yell, in the same way that "huh" is almost just an outbreath.

[1] https://www.smithsonianmag.com/science-nature/everybody-almo...

This route makes sense to me; I do think the deal has been getting worse and worse lately, and periodic renegotiation is required to keep things running properly.

My guesses - this is about (a) inequality reaching a breaking point, where non-elites are starting to reject the social contract, making elite signaling dangerous, and (b) that elites at that moment don't have enough justification for their elevated status (mystification, misrecognition, etc)

> Beyond that, airlines owe you nothing for your most valuable items. Most contracts of carriage specifically exempt from compensation things like antiques, art, books, documents, money, cameras, collectibles, electronics, or "fragile or perishable items." [1]

Also this is not a guaranteed payout, you may need to provide receipts for your things [2]. In general I would prefer to keep the contents of my bag vs receiving "up to" $2,500. We're mostly working people here. The time and money cost of travel disruptions, replacing stuff, navigating bureaucracy etc. is not low. And my stuff has emotional value.

I agree that this setup seems reasonable in a perfect world. But knowing how large, complex companies function, I feel it is unwise to create any kind of loop which rewards airlines as a function of luggage lost.

[1] https://www.frommers.com/tips/airfare/the-bottom-line-what-d...

[2] https://www.peopleclerk.com/post/airline-lost-delayed-luggag...

McDonald's is actually a good example, they did exactly this. IMO one reason their brand has been devalued is death by a thousand cost cuts. Each cut is imperceptible and seems like a win. But over several decades the net effect is a disaster. (IIRC Fast Food Nation documented some ideas about this process)

If airlines were prepared to accept X in costs from lost luggage before, and now get paid Y, they can now accept about X+Y in lost luggage costs. Meaning more lost luggage! After the system dynamics resettle, from a certain angle this is basically stealing: customers pay the same ticket price, but are paying some percentage more in expected lost luggage.

And on top of that, this is just considering the financial cost - what is actually happening is that more travelers are being deprived of their personal belongings. These are worth more to people than insurance payouts! It's surprising, but when you zoom out this seems downright awful.

[1]: https://en.wikipedia.org/wiki/Moral_hazard