hawtads

> Musk (who freely admits to distilling OpenAI's models) struggle to build a competitive 1T model (Grok 4.5) with massively more compute (Colossus-1 100-250K GPUs, Colossus-2 500K+ GPUs)

Are you sure they are using all of their compute on training? Didn't they rent out a ton to other AI companies?

> LLMs do not care about "elegance" the way human beings do, which is a big advantage.

It's just a matter of time before you can post train it for elegance too. Mathematical proofs in particular can be formally verified automatically which is a big advantage.

Isn't that rather self evident? If you are sampling from a particularly domain constrained vertical, how do you expect the censorship to transfer?

> The distillation data also did not contain any China-sensitive content.

This is a very big disclaimer.

It's like if I generate a dataset focusing exclusively on forestry and arboriculture obviously there won't be any useful censorship, or at least little that can be classified above a statistically significant threshold.

If you want to do a study on something more interesting and useful, do a piece on the various guardrail models of all the major LLM API providers. There are usually both input and output guardrails, and they tend to be almost-black boxes from the model routing point of view.

I don't see what the issue is. Isn't the entire point of LLMs that we can easily translate between languages? What better way to cross check the correctness of a proof than to run it on several different systems simultaneously.

This is just an API level change right? The model itself should be the same I think.

One of the best technical explanations I have heard in a long time.

For model routers, do they have to retrain the routing model every time a new LLM is released?

Not sure about OCR specifically, but the newer (past quarter) vision language models all have a lot of post training on detecting garbled text specifically. You can feed some of the old stable diffusion outputs into a modern model and they can figure out pretty reliably if the text gen is mangled. I think the feature is probably used as part of the RL for the image gen to correct for bad text rendering.

I think OP is referring to the initial sign up based on an invite link, not the database storage itself.

It depends whether you want to check the invite code against the invitee's email. It very much depends on the application. For most use cases, just verifying the code is fine and accept any email addresses that posses it.

Essentially ask yourself this, if you change the UI a bit and allow any user to sign up with any email, and the workspace joining is done via a code they enter after login, is that acceptable for your product design and security posture.

The word champ comes from the world of horses actually

https://www.npr.org/sections/memmos/2016/06/09/605796769/che...

*champing at the bit

If you are on react/next.js, defer the client side initialization until after your app has painted. PostHog (especially with their session recording feature) likes to initialize a little before the rest of your app loads through either their context provider or instrumentation.ts (if you are on next.js). It's generally insignificant except if you are building a B2C web app where the extra 100ms makes a difference in retention.

On the server side, queue all analytics call and run them after the main request completes (on next.js it would be within an after() function on the server side).

You can paste this comment into Claude and it should handle the refactor just fine.

Note with the changes in analytics scheduling, you will lose out on real time analytics in favor of better time to first load. So keep the trade-off in mind.

> What you're describing is a behavior tree: predefined logic, predefined responses, no learning, no inference, no model.

Stop calling everything AI, guys.

Depends on how pedantic you want to get, one could argue that regular expressions are AI too.

https://www.rand.org/content/dam/rand/pubs/research_memorand...

Regexes were invented for much higher order tasks (modeling neural networks) than just making find-and-replace easier.

He is losing a lot of information in that normalization pipeline (whole shell reduced/feature engineered into nothing but an outline). A CNN or something similar would be better and he can maybe get a better depth map of the mouth shape.

There is OpenCode and Pi, they both work pretty well

Why not use higher thinking effort?

Elenco is still around. I just bought a XP-720K linear power supply kit last week.

Have you tried Jetbrains Fleet? Their new editor isn't too bad.

We are working on making agentic ads and regulatory compliance scalable.

https://hawtads.com

Just launched the blog too

https://blog.hawtads.com/

The original Kleene Star Regex was invented to model neural networks. Have you tried throwing a transformer at the problem /s? Also O(n²) but at least you get hardware acceleration ¯\(ツ)/¯

Here's Kleene's Representation of Events in Nerve Nets and Finite Automata:

https://www.rand.org/content/dam/rand/pubs/research_memorand...

Ooh good find, thanks for the link. This will be my bedtime reading for this week :)

I am more familiar with Bayesian than frequentist stats, but given that they are mathematical equivalent, shouldn't frequentist stats have an answer to e.g. the loss function of a VAE? Or are generative machine learning inherently impossible to model for frequentist stats?

Though if you think about it, a diffusion model is somewhat (partially) frequentist.

I think it would be interesting if frequentist stats can come up with more generative models. Current high level generative machine learning all rely on Bayesian modeling.

Well, hope they reinforced the wings, that's a massive weak point for dusters.

50 knots rotation is perfectly fine for a plane that size. A Cessna Skyhawk is certified to rotate at 55 knots fully loaded (and since the stall speed is around 40knots, for specialty take-offs like soft fields it's much lower, 50knots is more than enough).

No, Claude on GitHub Copilot is billed at 3X the usage rate of the other models e.g. GPT-5.4 and you get an extremely truncated context window.

See https://models.dev for a comparison against the normal "vanilla" API.

We are building an agentic ad tech system optimized for real time and scale. The process of making an ad, from ideation to distribution, is traditionally exceptionally labor intensive. We are making it possible to target, design, and distribute ads at scale and in real time.

https://hawtads.com

I have been working on the next generation of Canva and Photoshop for highly regulated verticals where there are specific demands placed on the generation and edit flow.

https://hawtads.com

If you are a brand who needs to deploy advertisements at scale, don't hesitate to reach out.

It could have exceeded either its real context window size (or the artificially truncated one) and the dynamic summarization step failed to capture the important bits of information you wanted. Alternatively, the information might be stored in certain places in the context window where it failed to perform well in needle in haystack retrieval.

This is part of the reason why people use external data stores (e.g. vector databases, graph tools like Bead etc. in the hope of supplementing the agent's native context window and task management tools).

https://github.com/steveyegge/beads

The whole field is still in its infancy. Who knows, maybe in another update or two the problem might just be solved. It's not like needle in the haystack problems aren't differentiable (mathematically speaking).