NitpickLawyer
(opus <= 4.8 and gpt <= 5.5 couldn't match a PPO baseline network, and 5.6-sol and opus5 could barely match it. The 20x baseline improvement only comes from opus5.5 and sol > 6.1 / Astra)
I think you missed their joke :) the joke is that in europe (hence the Mediterranean) people often use "hazards on" as a justification for parking / stopping anywhere, especially where they aren't allowed to.
Yeah, but the majority is sane. There are only 3 vehicles in the park.
HA! Now that's cool. Not necessarily the improvements, those are ancient and probably overtrained benchmarks, but it's so cool that the model found a singular direction for this "loop bug". I imagine it could possibly do much the same thing in finding other "single directions" on a variety of topics (if allowed by guardrails). This is really really cool, especially for the mechinterp folks.
Eh, dunno. They've shown lots of promise "in the lab", on these narrow domains shown in papers, but AFAIU they're hard to train on some domains (i.e. language). Someone would have released something if it's that promising. Even RWKV had more models trained that showed real comparisons. Jepa seems like a theoretical toy, until they actually prove that it works in real world (hah!) models...
One possible use for these is to facilitate lab -> this -> real world for robot models. Basically better than lab / algorithmically designed environments and cheaper / easier to train on than real life. Something that "looks" like real life and can scale, to apply RL for robotics to.
A note that the only inference providers that hide thinking are the closed labs. Most (all, afaik) inference providers serving open models will send the clear thinking tokens always.
A malicious inference engine can already return anything it wants, it doesn't need constrained decoding on the LLM side. If your client blindly runs commands in an unprotected environment, you're already pwned. They don't even need to run an LLM, they could just randomly add "tool_response:{'execute':'rm -fR *'} and that'd be it...
Yes, this is what constrained decoding means. You constrain the model to only output from a set of "allowed" tokens (a.k.a. grammars).
So, if you use tool use with constrained decoding, you can be certain that the model will always choose from a list of approved tools (i.e. if tools=['read','write'], it cannot output 'execute', even if in other sessions it could, if execute was in the approved tools).
The tricky problem is that certain models work / have been trained in certain ways. And constraining them can lead to "invisible" problems. Constrained decoding only guarantees perfect syntax, not perfect semantics (i.e. it could write nonsense in the correct file).
> Could something like that also be used in other contexts?
Sure. You can build pretty cool grammars and constrain the model to do cool stuff, akin to "tool use" but more free hand. You could, for example, make it always emit a 10 point plan. For any prompt, it will give you a fixed syntax 10 point plan, always. And the plan will be constant with a plan_title and 10{plan_item, item_text}. But, as above, there's no guarantee that the individual items in the plan would make any sense... (i.e. step7 in a pasta making recipe would be "add glue").
Yeah, there's another paragraph:
> Identifying and removing cyber-related data. During development, we observed vulnerability research performance improving steadily as post-training scaled. Tracing the source of this trend, we found that cyber-related content was embedded within the general mid-training corpus, and cyber-related tasks were present in the RL environment mix — a natural byproduct of training on broad data. In preliminary experiments on internal models, we found that without data removal at either the mid-training or the RL stage, downstream cyber capability would increase substantially. These findings gave us a concrete lever for control. For the open-weight release of K3.1 Pro, we identified and removed cyber-related data from both training stages, while retaining the general-purpose coding tasks needed for model usefulness. We will continue to apply appropriate controls at every training stage, of which cyber-related data identification and restriction is one example.
(emphasis mine)
It's clear that cyber is a touchy subject right now. But it's also clear that RL works on it, because it's highly verifiable. So smaller models are going to become very interesting, because they can be fine-tuned later to cover this field.
This is absurd, and it shows that you have absolutely 0 idea of how any of this works. Finding a bug and delivering a working RCE 11 days later is absolutely not "sitting on it". This is how this works. You need time to test, exploit, confirm, and so on. Sorry, but just stop being so negative. This can't be healthy for you.
> 1 August 2026 [...] Ryan Austin finds the RCE paths, tweets a cryptographic hash on 5 August, and submits the bug and working RCE proof-of-concept to PHP between 11 August and 14 August.
> 19–24 August 2026: Wiz Research independently collides with the same bug family using Wiz Atlas, Gemini 4 Argon, and GPT-5.6 Sol, reporting its variants on 24 August.
I get some skepticism about marketing copy, but your comment is blatantly wrong and needlessly rude.
- some countries require saving "the last known address"
- others require keeping logs about "the last known connection point" (IP)
- in some countries, if the user was a paying user, you need to retain payment details / invoices / etc for years
- some countries you can receive subpoenas that mandate what data should be retained
All these things need to be checked before deleting an account. Having a "manual" deletion endpoint would be more troubling.
As an aside, early in the GDPA implementation phase, there was coverage of many companies implementing it poorly at best. There are articles detailing how you could delete a user's account by sending the "required info" (e-mail, ssn) from another e-mail and they'd delete the account / send over data. Would you be comfortable with someone being able to delete your account like that?
Why do you think that? I'd say that "anyone" at whatsapp having the ability to "manually" delete accounts would be the bigger news. In large corps, no one should have that capability. Anything dealing with user deletion should go through a well established, legal approved process.
I'm pretty sure someone will soon call Linus a sell-out. The hate / denial / radicalisation is just sad at this point... Do, or don't, but at least don't hate the others. Be and let be. No need to call other people names just because you don't like their sand castle. There's plenty of sand on the beach.
> Then they set their identical records spinning at the exact same moment. On each end, the voices of these two world leaders are threaded through the SIGSALY terminal, mixed with a layer of noise, sent across the ocean via shortwave radio, then re-synthesized in a reverse process.
> Anyone attempting to listen in on the calls would just hear a buzzing sound, which is how the system earned one of its nicknames: “The Green Hornet,”
> At the end of each conversation, the records were destroyed to protect against potential future decryption.
I wonder if with todays tech we could decrypt this. This sounds really close to a one-time pad (if the conversation was kept shorter than each record) so .. probably not? Unless their random noise wasn't as random, or if the mixing of random + voice would leave other tell-tale signs?
All of the above, to the extent that they can be measured from a simple to verify end condition (stats, win rates, etc). Adaptability, ability to understand and progress on a task, and so on.
While the 2nd half of the book is definitely still sci-fi territory, the first part is slowly getting into the plausible tech realm, 12 years after the book was written. We now have a much stronger space industry, and probably could match or exceed the payload to orbit in 2 years from the book (if everyone would work towards the same goal, which is still a bit of fantasy but hey...)
Impressive capabilities shown on benchmarks don't really make a difference if you can't use them. I'll take a lower chance of success over constant downgrading and refusals. At least, as the author put it, you get to enjoy the token gamble exercise.
Partially agree and partially disagree. Qwen 3.8 next (120b + 50b ngrams) is at the upper end of a local model, and serving it at good speeds (esp for pp) is not cheap. So no, oAI are not particularly scared that everyone will suddenly start using local models of this size / quality.
I do agree that having these models in the open is great pressure on the price point of "small" API models (like luna, haiku, etc). And that's good news for us consumers and for the field in general - more competition is always better.
And on the cost estimate, the true power of local models is the guarantee of running them on prem w/o any data going out. Price wise it is still cheaper to go API, for small teams.
> write a more generalized SDR application for this hardware
Yeahhh, no. Sorry but the entry barrier is now extremely lowered, especially in effort.
~20 years ago this would have been literally sci-fi. Back then just getting an USRP to work with gnu-radio, then finding the proper hardware external clock, then doing lots of work on reading / trying stuff / writing scripts etc. took a ton of effort and time. Having the ability to throw tokens at it and have at least a starting point is amazing, IMO.
Don't fall for the "marketing hype" antihype. RSI is a trendy term today, and lots of people discuss it, but it's nothing new. We've basically had forms of RSI since the start of the LLM craze. Previous models have been used to improve next versions for ~3 years now. Just because humans used to guide / select / decide when and how to apply those models, doesn't make it less recursive or self or improvement. We're now seeing what happens when we remove the humans from some of the tasks, but the basics are still there. Use prev model to make next model better.
Some people saw "the spark" before LLMs were mainstream. We had code autocomplete models based on GPT2 that ran on your machine and provided line-based autocomplete. Then we had gpt3.5 (chatgpt launch) where it almost looked like it could write python, then we had gpt4 and saw the first glimpses of actual code writing, then opus4 / gpt5, and today we have sol/fable etc. At every step we had people in this field write the same kinds of takes. And at every step the tech evolved, improved, and got better. To a point where the vast majority learned to accept it, and the denialists (there are still some) are now in a minority.
And the acceptance phase was not "we are now useless as SWE", nor was it "hah, it's only for the juniors, we are safe. It was, mainly, "we can now work at a higher level of abstraction, and use our experience to guide these tools and work faster / broader / deeper in a topic. Basically understand the tools, learn their pros and cons, and adapt at using them.
Also part of it could be that the message seems to complain about the situation without providing alternatives. What are the alternatives? What are they proposing? "Please don't do this" does seem like gate keeping if you don't offer an alternative. What exactly do they want these labs to do? Don't try to solve the problems? (stop the count?) Don't publish the results? It's not clear to me.
I guess a lot of people are unclear on the messaging. Is this purely ideological, or is it something more? (It doesn't help that AHM also has a section where they highlight their members who pledge to do research "without AI assistance". Whatever that means. Yeah, that's cute, but can signal denial, which again we've already seen in SWE, and hopefully most people have dismissed it already) Are these statements part of the 7 stages of grief, or is there more? When taken at face value, the statements seem rushed, unprepared and "in denial". At least when compared to writing on the same topic from other people in the field (Tao et all).