Amekedl
Many such cases are benign in the sense they in practice are simply the model basically "git clone"-ing the code they need to complete the implementation.
The amount of recall and synthetic data slop makes it imho very hard to actually get the models to really do something more novel (where they fall apart more often)
I can't help but feel how this is all industry-planted, and I wouldn't argue its too far fetched.
It's always been the last 30% where it just fails.
Guess it is 870M to 0
I mean, claude already does it for you, constantly even
Now just deno install everything and even desktop works amazingly well.
Hard though from the idea initially too.
The main stack is reasonably web, as broad as that is
Enough stuff can happen, software use itself might change, and that could really cause anything. "What will we do with all the gpus" might become a question if for a magnitude of tech and reasons leaked-opus-9 runs on a macbook m6 or 7
Corporations, like Microsoft, offering 5 grand to that in comparison seems just lackluster.
Relationship and whatnot, pay that teenager 100k is what I'd say, really.
Offering him a job, that would be cool too.
In any case, I'm hoping stuff like this really remains a "once in a lifetime" opportunity for pen-testers and whatnot.
Having the planet hacked twice a day makes for a stressful future for everyone.
Tweaking everything, because any agent/model can only "interpolate" so much "resolution and detail" out of your written prompt.
This isn't really a problem, this is the nature of work, and hence I feel this product only accomplishes two things:
1. users spending more on inference
2. creating busywork with less direction than other surfaces such as a IDE, which is hard to review and will often lack meaning.
But even when starting that, I was aware such patterns and remote access, yeah, this just has C&C botnet capabilities too.
It's not that deep, but there certainly is overlap with malware-esque patterns.
Especially with the hardware stuff, this is plainly put unachievable by many IT departments.
The fact that models vastly outpaced their harness and permission systems - I wouldn't dare to doubt this fact.
Claude Code on auto is still rolling a dice with its sonnet classifier - whether that IaC action I told it and explicitly stated multiple times it is permitted and authorized to run - yet it always randomly allows or denies it.
Therefore this is totally still a unsolved, perhaps unsolvable problem.
Realistically; anyone paying for llm access (anthropic, openai, gemini), is getting their access, and a service provided billed by tokens, subscription, whatever.
All the efficiency gains, which publications like deepseek v4.1 flash seriously frontload like it is their most important topic to have accomplished improvements on without diminishing performance too much - now this is a thing anthropic and anyone else also cares about, but for different reasons.
American "providers" with closed models are setting their token pricing somewhat arbitrarily, which is fine: it means more profit, and pretraining and RL experimentation is super important and expensive.
They (closed model providers) have very likely super optimized inference too, just like deepseek, but it's not at all something that any customer really has to care about - they just want the service to be as cheap and great as possible.
Sometimes it is really akin to a git clone, although it was a LLM request.
This rewarding expertise is somewhat wishful thinking.
At the end of a day, it feels and is more like gambling, even with the recommended expertise and a good approach, don't delude yourself you're simply pulling the lever too, as any novice.
FYI, I prefer discussion over being mad
The "anti LLM" or "no AI" clause is really more a "covering all ass" clause.
If you want to improve Forgejo and have that AI PR; you better clean it up of all traces - resulting in a PR that is likely neat and going to be merged.
I see no issue with that. It simply enforces a pace and process that is very befitting and in my opinion mandatory for such a central system they're maintaining.
1. "Grokking" was shown on 4-digit modular arithmetic with a 1-layer transformer; this article extrapolates it to AGI and a $10B training run with exactly zero intermediate evidence.
2. The "small dataset" is 25 trillion tokens - literally the size of current frontier training sets - but calling it small sounds revolutionary.
3. BabyLM has spent 4 years failing to produce grokking on constrained data; the paper gets a footnote saying "those models were too small," which is unfalsifiable until someone burns $10B.
4. Chain-of-thought is already empirically required for frontier performance - it's expensive, bizarre, and nobody predicted it - yet somehow we're supposed to bet the farm on a phenomenon that has never scaled past arithmetic. We need that data, even if it is just "Actually, ..."
5. If you want to chase "recurrent depth", loop transformers rumored in Mythos/Fable are at least grounded in actual engineering; grokking-at-scale is just vibes ai bro science.
More data is and will always be the answer. Why are all labs distilling from each other?
As if it really would work like that. The noise added by the verbosity alone is not taken care of enough, and this entire thing belongs on the great pile of ai vaporware.