Does the model quality become degraded in other ways?

no because we freeze the main weights and only train our added weights.

It's impossible to add weights which degrade the quality?

Not if the new weights don't influence the old outputs.

Aha! I thought this was feeding the confidence info back to Gemma for some reason.

Do consider the incentives of those developing the model that made those predictions. Afaict, it was not selected for purpose other than testimony.

How do you get personalization without surveillance?

This is my bad. We've heard a few people complain that the word "personalization" has been adopted by the ad industry and it makes people shiver when reading it. In our case it means "personal search customization", meaning you can customize the experience as you prefer (UI, domain rewrites, search providers, boosts, etc.). We don't log or track any search queries. I've fixed this in a couple of places but it seems I missed some others, so I'll be fixing it promptly.

Meanwhile, I hope that answers your question? Let me know if you'd like some further clarification.

I suspect idiomatic TypeScript or idiomatic F# are both way better solutions in the real world than abstruse Typescript emulating idiomatic F#.

Possibly, but I think what we wish for is a language with a nominal type system that lets you switch to structural typing when needed.

Luckily, F# has type providers, which lets the compiler construct nominal types based on the structure of real data (like json, xml or any format you want), saving you from the effort of building wrapper types by hand.

If a good SWE is $150/hour, does the model cost actually matter? Surely you'd be willing to spend $10/hour to make that SWE 20% more productive? The model cost is still much less than the salary.

I don’t think any engineers who cost $150/hr are having their productivity moved by 20% depending on a $10/hr gap between models on or near the frontier.

Most of the gains right now come from tooling and process and any big post 2025 language model. The specific model isn’t that important right now.

Exactly. And being able to choose your own tools is much more valuable than having a tiny bit better model.

With Claude Code Ultrathink, I used 3 million tokens in 20 minutes. At API prices, that would be around 30$. So 90$/h. Model cost is not that much lower.

x40hrs/week * 50 weeks = $180k

Congrats, now you’re paying an engineer’s salary to make your engineer at best 20% more productive.

Better to hire another engineer, or two jrs, and build up your in house talent.

Only you get things done lost faster and don't need to pay entire years salary?

How so?

If you’re doing a one-off project, sure. But if you’re coding like this full time, you’re paying the year’s salary anyway.

And faster? Not so sure about that. Sure they can write code faster, but writing code is a small part of building something.

That's a good point.. Writing code vs building

except this is way more than an engineering salary. At least in Europe.

I’m sure there are engineers making $180k usd / year in the eu. Maybe it’s unusual, but hey, now you can cancel your claude subscription and hire a really good engineer

But SOTA models used liberally at API pricing is a lot more than $10/hour. You can probably burn $100+/hour with just a single agent, and probably thousands when running agents programmatically, e.g. workflows.

They use a lightweight adapter to silently degrade the performance. Usually these adaptors are made to improve the performance for a given domain/task.

3 blue 1 brown has a great visual introduction to transformers, the heart of LLMs.

It's chapter 5. Start at chapter 1 if you want more background on neural nets and backprop.

https://youtu.be/wjZofJX0v4M?si=HFXbrB-5cArprGaU

"The reasoning is the weights."

The reasoning is in a process that uses the weights.

Sorting algorithms are just bytes. Those bytes don't sort by themselves. They do instruct a computer on how to sort though.

There is some recent work on modularizing knowledge in LLMs.

https://arxiv.org/html/2605.06663v1

It might be possible to train a big generalist that is a composition of modules, some of which can be dropped dynamically at inference time, depending on the prompt.

Cool. Thanks for sharing. I am thinking about creating a series of smaller models for specific purposes and then orchestrating them so they mirror the human brain which is a bunch of subsystems that give multiple opinions about the same stimulus

Interesting direction. I’ve also been thinking about modular / subsystem-based approaches for specialized tasks in small AI systems.

Is every American tax payer morally compromised?

Yes ;)

I agree with the intent of your rhetorical question, so I'm jesting with you. I'm justifying my "yes" with the hopefully humorous distraction that every person, including American taxpayers, has at some point made a nonsustainable/selfish (my definition of immoral) decision.