David Dworetzky 🇺🇲

AI tools/infra Into AI and Healthtech. Jiu Jitsu🥋. Dad x 3 👦. Classical guitarist. AI agents and Privacy / Legal Tech.

[David Dworetzky 🇺🇲 @DavidDworetzky]

A bit of life and machine learning philosophy - everyone is solving an optimization problem with a different reward function.

[David Dworetzky 🇺🇲 @DavidDworetzky]

I feel for the Engineers working at GitHub.

The execs on the other hand - guys, get it together.

[David Dworetzky 🇺🇲 @DavidDworetzky]

@yacineMTB This tweet took place in Minecraft.

[David Dworetzky 🇺🇲 @DavidDworetzky]

Who’s running a single long running session for everything with their agents - like a long running perpetual stew? 🍜

[David Dworetzky 🇺🇲 @DavidDworetzky]

I’m surprised at how many agent harnesses view voice as an afterthought. Voice and BCIs are likely closer to a first class input form in the future.

[David Dworetzky 🇺🇲 @DavidDworetzky]

Felonybench - how many felonies your advanced AI commits while escaping its sandbox.

[David Dworetzky 🇺🇲 @DavidDworetzky]

@thdxr Inshallallm

[David Dworetzky 🇺🇲 @DavidDworetzky]

Excited about buzz and decentralized agents.

[David Dworetzky 🇺🇲 @DavidDworetzky]

and suddenly thousands of instances of fable cried out Eureka! as a Jacobian counter example was verified again and again…

[David Dworetzky 🇺🇲 @DavidDworetzky]

So K3 is open weights - it changes the dynamic quite a bit for the labs. Clearly in a way that is very pro consumer surplus for all of us - and better for research.

Also I think it is moving AI behavior in the direction of being pro-agency and tool shaped.

[David Dworetzky 🇺🇲 @DavidDworetzky]

Excited to hook in K3 to some harnesses this weekend.

[andrew chen @andrewchen]

This is what I want to say:

Been playing w GLM 5.2 for the past week and it’s legit. If I can get this in a box sitting on my desk, I’m not sure I need much else

But this is probably what’s actually true:

Can imagine a guy running on a Pentium PC for the first time and being like, wow, I don’t need anything else. But actually we invent all sorts of cool new things we can use compute. So each frontier/adjacent LLM is cool but we’ll always want to use the next, better one

[David Dworetzky 🇺🇲 @DavidDworetzky]

1. Fable is really good.
2. Yes you still have to read code.
3. You probably need better test & loop infra than ever.

[Delip Rao e/σ @deliprao]

Dario says, “AI fundamentally changes what is possible with surveillance”, and I 100% agree — it’s also why you should use local models that you control and limit your data exhausts to closed AI companies.

[Shannon Sands @max_paperclips]

ok fine Fable is a good model. like....really good. insanely good in fact. "clean up this old PoC and the research I had from the last couple years plinking away at the problem........oh, it's all working now wow. and you improved that thing. ummmm ok, I wasn't prepared for that to now be working wtf do I do next" good

[David Dworetzky 🇺🇲 @DavidDworetzky]

I talk about this more in tokens are not enough.

https://t.co/Pcg1cWxzOf

[David Dworetzky 🇺🇲 @DavidDworetzky]

Engineering is evolving faster than it ever has.

I think code and craft will always matter - even if we’re all salespeople and product managers at the same time.

I think Anthropic’s product is great - but there’s an aspect of how Boris markets Claude that rubs me the wrong way

[David Dworetzky 🇺🇲 @DavidDworetzky]

Law is the architecture of freedom. Happy 250th birthday America! 🇺🇸

[David Dworetzky 🇺🇲 @DavidDworetzky]

Local models and Cloud inference will both always have their place.

Hard to see without step change differences in energy density or compact capabilities that you won’t want to serve things out of a monster cluster.

But there are reasons of privacy and embodiment that local models will win on.

[eric provencher @pvncher]

The job is different, but you just can’t expect to have a scale product without technical people steering and inserting their taste into the code shipped into production.

The software engineers at OpenAI are busier than ever, even if the amount of code written by hand is near 0

[Arcee.ai @arcee_ai]

Today, we’re releasing the first weights from Trinity Large, our first frontier-scale model in the Trinity MoE family.