431–440 of 626 posts
Earlier quoted context omitted.
Nvidia has made a lot of very suspicious circular funding deals. I suspect we’ll find fraud when the bubble bursts yes.
You're the first person I hear claiming NVIDIA is hiding billions of dollars in debt and fake profits, never mind at the scale of Enron. Bold claim!
Earlier quoted context omitted.
Can’t this be extended quite far? Use a cerebras-served model, use verification techniques to generate and solve millions of problems and then use that as training?
This isn't latency bound, it is trivially parallelize. So you want to run it on the most efficient compute you have, not the fastest.
Earlier quoted context omitted.
I’m in the Claude code harness for everything boat too. What are the alternatives?
I’ve tried a bunch of them, and I seriously do not understand these recommendations. It was a rough road and a steep hill, but right now CC is absolutely the best harness on the market, as for me, whatever top tier model is under the hood (mostly, some of them, like DeepSeek, don’t fit CC at all).
I bought $18 GLM official subscription yesterday (5.2, but new model version was already leaking on some docs), set it up with Claude Code harness... and I’ve bumped to $80 plan almost immediately. It’s the first model that agreed on a proper security research (red team scenario), executed it seamlessly, including 0-days in WP plugins, RCE, 6.8 kernel exploit adaptation, etc - while playing against another GLM agent…
“ Open Source: We will release the weights in two weeks after launch, once safety evaluation and hardening are complete.” Cybersecurity capability might be nerfed
Earlier quoted context omitted.
I’m in the Claude code harness for everything boat too. What are the alternatives?
Piggybacking on this thread to ask my question: What are alternatives that are multiplayer (team oriented) by default? For example, I want my team to see all my sessions easily, vise versa. another way of stating: all the agents are running in a container that that any member of the team can view and interact with.
Earlier quoted context omitted.
I’ve tried a bunch of them, and I seriously do not understand these recommendations. It was a rough road and a steep hill, but right now CC is absolutely the best harness on the market, as for me, whatever top tier model is under the hood (mostly, some of them, like DeepSeek, don’t fit CC at all).
Funnily enough, I would say almost the opposite. CC’s feature set is basically table stakes for an agent these days (does it have ACP yet? Very close to behind table stakes if not) and it has a lot of bloat powering that. IMO part of it is that the underlying LLMs have gotten better enough that harnesses feel better even if they haven’t changed. I have a toy harness that barely implements the features you’d expect an…
Edit: lol, I don't think ACP is even actively developed anymore. It seems to have been merged into another seemingly pointless standard with an even worse name, A2A. [0]
People familiar with the topic, how will models continue to get better? Post training it seems? Labs have already used up internet-scale data, so are there any limits to architecture improvements and post training or can we expect this trend to continue? ByteDance is training a 10T-parameter model. Here, GLM 5.3 outperforms models 3-4x its size of roughly 700B, so parameter count doesn’t seem to be a direct correlati…
> Labs have already used up internet-scale data Despite this being the topic du jour of 2025, it was never true. Most of the "we've hit a wall with data" came from communicators / media and not researchers. It got popular because negativity sells. It's a false premise for a number of reasons: a) Data curation is as important, if not more important than bulk data. Models becoming better at classification leads to bett…
Big models are going to "tap out" on non verifiable fields within ~2 years, just because the pool of experts able to reinforce the models is going to get very small, and as the nuances get finer, the signal from reinforcement is going to get progressively less aligned with the intent. Math and code will be mostly tapped out in that time frame as well, even though we can technically scale them "infinitely," just because the cost benefit won't line up. At that point, most RL will be "gyms" with games that are designed to model designated valuable economic activity.
In the next few years, we'll get small domain specific distillates that are ridiculously smart in their domain (imagine if Qwen 3.X 27B went super saiyan), and even frontier labs will be routing to experts/orchestrating because the cost to serve/TPS difference is huge. They'll still train the god models for PR/marketing, c-suite use and distillation, but using them for day to day work would be like making houseware out of solid gold.
Earlier quoted context omitted.
> Anyone using them now? You're gonna have a had time getting straight answer to that out of the internet. There are now 4 different flavours of the Max plan floating around (Legacy V1, Legacy V2, New plans, and the current credit ones). And on top of that they have peak times. So ~8 scenarios, 24 in total across all feedback for their coding plans. So when someone tells you they're having a good time on a GLM coding…
I have V1 Max, and I think they throttled me for using it too much. I was maybe abusing it, by sending out 8 or 16 review agents at a time. I haven't tried it in a few months, but it went from amazing to unusable really fast.
Earlier quoted context omitted.
> I understand the argument of "people are not actively looking", but isn't the cost for such a scan getting lower by the week, and Anthropic's Project Glasswing is supposed to find them quite a while ago? You have to consider that having an LLM scan for vulnerabilities is hardly infallible. It is a search guided by heuristics and given a large enough codebase, it is unlikely to identify all vulnerabilities. Personal…
> [A]ll of them found some that the others hadn't discovered. Now, correctness issues aren't the same as vulnerabilities, but the same principle about using heuristics to find defects applies. This makes perfect sense, but that conflicts with the impression put forward by Anthropic and OpenAI (in particular) that they alone occupy 'frontier model' spots. Frontier models should large dominate their competitors on a ca…
Not necessarily. Even near the frontier, we don't really have a total ordering of capabilities, but a partial order. And even frontier models make plenty of mistakes. Combined with the randomness inherent in searching large codebases for vulnerabilities or correctness issues, it is entirely plausible that even much weaker models (and GLM-5.2 isn't even weak) can stumble upon issues that stronger models missed.
My current hypothesis – for which I have only limited evidence, unfortunately – is that it is better to have multiple reasonably powerful (but not necessarily frontier) models looking for issues than just one very powerful one. And even then you're likely to miss out on some issues.
