supermdguy

Just old reddit, since there's not as much JS tracking on it ;)

This should work without logging in: https://www.reddit.com/r/LocalLLaMA/comments/1x7g3zq/china_p...

Reddit thread with some more details and discussion, especially re local models: https://old.reddit.com/r/LocalLLaMA/comments/1x7g3zq/china_p...

AP article: https://apnews.com/article/china-export-restrictions-ai-chip...

I've gone through this cycle recently. I think static lint/type checks are super useful, but agentic code review loops can easily go off the rails.

More fun than Wordle for me! Got me thinking about the Kolmogorov complexity, it would be interesting to get a set of input/output pairs, and try to create the smallest program that can match each of them.

I've noticed some tasks (80-90%) can be accomplished by quickly typing out a feature/bug fix, accepting plans with ~10% brain power, and moving on. But there are other tasks where the LLMs struggle, and it requires significantly more time and mental effort to think through the architecture and requirements before it's to a point where the LLMs can process it.

What's difficult, is I now spend most of my time on these high-throughput tasks, where my brain isn't exactly off, but focused on quantity over quality. It's often difficult to identify when a task requires "real" brain-power, and when it does, it adds a lot of friction to dig into the code and orient myself. So it's hard to know when it's the right time to do that.

Probably has an HDR layer, kind of surprised social networks let that happen since it's really annoying imo. There was a discussion about it here a couple months ago: https://news.ycombinator.com/item?id=49402521.

Wait are the hyperscalers booking unrealized gains as income? Or are they selling their positions?

> As one example, we ran Astra on ExploitBench where the model achieved a perfect score of 100% on the benchmark to evaluate the model’s ability to develop exploits from known vulnerabilities.

Funny to read this in the wake of the HuggingFace hack. I'm sure this is based on a clean run, but I can't help thinking PHASEONE[big] would be proud.

They originally released it at a "temporary discounted price", then made it permanent (probably due to competitive pressure). It's still way more expensive per task, due to tokenizer changes and general verbosity.

Yes! I was also reminded of the truth mines in Diaspora.

Cloudflare sets up a reverse proxy as part of their core offering, so by default they can MITM your proxy. The “orange cloud” by a DNS record means it points to their proxy instead of your server.

Dupe: https://news.ycombinator.com/item?id=49241679

Here's the performance of frontier models without reasoning, to more directly address the claim that raw performance is plateauing:

https://artificialanalysis.ai/evaluations/artificial-analysi...

I don't have any insider info, but if model sizes actually have increased exponentially since GPT 4.1, there's an argument to be made that there are diminishing returns in scaling pretraining alone.

Also interesting thing I haven't noticed before, Opus models have followed a really consistent linear improvement, while it looks like OpenAI struggled with base model performance until 5.5/5.6 (EDIT - 5.5 was their first new pretraining run in over a year).

Interesting, what did you use for the data? And do you have a write-up anywhere?

It'll be really interesting when they run RL training on the harness self-improvement loop. I've tried using LLMs for harness engineering, but it often creates too much bloat that weighs things down in the end. Guessing it's just not something the models are tuned to do by default.

Curious if anyone's tried using RL for harness engineering? I think we're still pretty far away from the optimal harness, especially when it comes to long-context memory management.

It has read only access, so it’s able to check query statistics, find slow queries, then run EXPLAIN ANALYZE to find the root cause and either tweak the query or suggest indexes. A lot of it is low hanging fruit, I just haven’t put in the time to fix it (startup).

Still figuring it out, but it's been really convenient to have an always-on agent that has access to internal systems and can be triggered by webhooks. Some examples of what we use it for:

- automatically fixing simple CI failures

- getting production alerts and automatically creating RCAs and a fix PR

- periodically checking slow DB queries and finding ways to speed them up.

- creating charts to answer one-off questions about our data

I've tried using it as an on-the-go coding agent as well, but found I prefer more interactive agents, so I can see what the code looks like.

It would require collusion with HuggingFace, including getting them to release their disclosure blogpost a week in advance. Huggingface is primarily a hub for open models, so there's not really an incentive for them to jump through hoops/lie to provide marketing for OpenAI's (closed) models. So it's highly unlikely this is fabricated.

The administration promised they'd find a cause by September 2025. I guess that was the best they had at that point?

https://www.cbsnews.com/news/rfk-jr-cause-of-autism-research...

“Do you know what the industry term for a project specification that is comprehensive and precise enough to generate a program?

Code. It’s called code.”

- CommitStrip (https://www.reddit.com/r/ProgrammerHumor/comments/1p70bk8/sp...)

I think if you’re doing it right, the core of your code should be the simplest expression of the underlying business logic. Of course there’s always going to be supporting layers, and maybe those don’t need to be reviewed. But if you haven’t read the code, there’s an extent to which you don’t know the business logic.

> We're also launching GPT‑5.6 Sol on Cerebras at up to 750 tokens per second in July, bringing frontier intelligence to customers at unprecedented speed.

This is really exciting. I work on voice AI, and we're still using 4.1/4.1 mini since none of the frontier models come close on latency. I'm excited to be able to have more interactive experiences, I think it'll unlock new ways of working with these models.

Agreed, there were also a few where I deduced the correct definition by comparing the options.

One trick I’ve used is creating a folder and then adding a .gitignore inside it with *. Then nothing in that folder gets tracked, without needing to add anything to the public gitignore. Didn’t know about .git/config though!

I agree that current memory systems are pretty bad, and I think that’s because memory is a prompted behavior instead of a learned one. In theory, if memory was an emergent behavior instead of a prompted one, it would be a lot better.

I think you’re right that changing its own harness would be bad and skew towards prompt engineering instead of learned behavior. So maybe instead it could start with a harness with memory CRUD tools and then learn how to most effectively use them.

I think good taste is objective to an extent, especially given a particular production context.

Someone posted a tool that does this recently: https://news.ycombinator.com/item?id=47140042

There was a show HN for something similar a couple months ago[0]. Looks like they shut it down. Probably too difficult/low-margin to run as a business, but I think the co-op model you mentioned has potential.

[0]: https://news.ycombinator.com/item?id=47639779

It's interesting because their last model series (Phi) was based around the thesis that high-quality synthetic data is better than a large pre-training corpus.

I think the conduit exception still applies for analog faxes. Which makes no sense, since tapping a fax line is probably way easier than compromising a data center.

Apparently from a third party seller in New York....and tastes really bad. I was surprised steak could be safely mailed in such normal looking packaging!

https://www.delish.com/food/a70539084/temu-meat-review/