We don't know what the prompts were though. Terence Tao showed of some of his, and they are indeed huge, with lots of context, things that might work, things he knows don't etc. What we do know about this, is that 10k agents worked together on this. The initial prompt, while probably still very relevant, would get diluted over time. We simply don't know the level of inolvement of humans in reaching this proof.
Agreed, there's a couple of bad behavioral shifts I noted for me when doing agentic coding:
- Doing too much at the same time, it is super exhausting, and I get nervous and stressed
- Sycophancy makes me overestimate myself
- When my stress meets external pressure, quality goes down. The little voice of 'surely the AI got it right' can lead to defects and code quality going down.
Except for doing some manual coding, I've started doing a couple of other things that are probably just healthy regardless of using AI heavily in day to day life:
- Do literally nothing for at least half an hour. Read something written by humans, even better if it is fiction.
- Do something that makes you feel stupid. And don't use AI to help you through it. (That could be Leetcode, but for me it is learning about statistics)
- Slow down when you feel like things are moving too fast. By far the hardest thing, but the moment I realize I'm trying to rush for a deadline, I'm going out for a walk. Realistically, I'm moving 3-8x faster than before, this little delay is nothing, and keeps me sane and quality on a good level.
Interesting way to phrase that. The easy way to solve rape is to castrate all men. We can save some sperm in case of wanted pregnancy, but since the vast majority of violent and sexual abusive crime is committed by men, it would be the most rational solution to the problem. Eevryone would be safer, even us castrates.
I guess we bave or should have as much control over immigration/asylum policy as we have over the diceroll of sex determination. Interesting gamble, cant see anything wrong with that. Also, why stop at males, females rape too so in the spirit of equity they should undergo the same prophylactic castration and that will be that
At least now riders are being added to applications and agreements that dispose of any cultural argument or litigatory posture that various misbehaviors were not understood to be barred prior to admission or status being granted.
I combine it with OpenCode + a small OpenRouter budget. GlM 5.3 Flash, Luna, Gemini 3.7 Flash and a couple of other very cheap models are sufficient for a lot of tasks if put on the right road. Telling Astra to debug an issue should be a last resort, it would blow through 10% of the session limit, but cost like 10c on one of the open models. Same goes for exploration, documentation, configuration and smaller features. These models are generally good enough, and you can still do a review with a strong model for a fraction of the cost.
The comment reads like flesh-generated language to me, but I might be fooled. The breadths of pro-AI sentiment is possibly astroturphed, but there's also an argument for an enthusiastic minority of people to write a lot more comments.
As the person above said, AI has significantly increased my ability to execute on ideas. I always liked computers, always liked building things, but was never that great at coding, and was never that good at going super deep into one topic. Instead, I have broad knowledge of a lot of things like product design, requirements engineering, devops, security.
I work with a fully agentic flow, but I would argue very seriously; My job has not gotten easier, I am doing at least as much hard thinking as before. From assisted RE over design (TDD focused Spec), implementation by agents, review by agents with partial human oversight, I have a speedup of maybe 50-100%.
More importantly, I can do things I couldn't do before. And when it comes to performance and defect-density, it is comparable to very senior people I couldn't touch before. I do agree that the code doesn't look like a human would write it - too abstracted, sometimes convoluted, often way too dense. But it isn't worse code, and if you accept that no one has to read that code ever again, then it is good. Agents are able to grok it just fine.
> When we offload decision-making, we become unaware of the trade-offs.
My biggest issue with the new agentic workflow is that I'm constantly asked to make decisions, and it is not always immediately obvious how important they are. Like, yes, I see the people who just write 'implement feature x' and then call it a day. These people were lazy und uncreative to begin with, and will continue to cognitively decline with AI. But if you use it seriously, you're in a constant state of doing requirements engineering, weighing trade-offs, and making architecture decisions.
> I wonder what the 2026 equivalent of going to California is now, if such a thing even exists or can exist.
I think this becomes increasingly difficult, because the timelines become so short with almost-instant information desimination, higher inequality and increased mobility for everyone. A place or event can go from underground -> awesome -> gentrified -> 'dead' in a manner of a couple of years now.
I live in a neighborhood that is currently being gentrified; 10 years ago no non-local wanted to live here, 5 years ago, it was the place to be, today young/hip people are beginning to move elsewhere again. Newly rented-out flats are now thrice as expensive as 2020, and you can see in real time how bars and restaurants are replaced with more expensive alternatives. When I compare this with other neighborhoods in Berlin (eg. Bergmannkiez or Kollwitzkiez), the timeline has roughly halfed.
Strongly disagree on Django, unless it is something you are already at a senior level at pre-AI. I've had the displeasure of cleaning up multiple Django backends lately, and the combination of standard fail-open, weak validation, mediocre ORM, and bad testing frameworks lead to issues I've simply not had when managing agents doing Go, Java or Rust.
Go probably wins as a matter of trade-offs for pure productivity (speed, reliability, ease of refactor), but Java Spring Boot works excellently if you're willing to take the plunge (pretty steep learning curve), and the ecosystem really lends itself to building more complex stuff that holds for a while.
Rust is fun because it is an amazing multi-faceted language where agents will constantly deliver you working yet surprising implementations you'll have to quintuple check, and sometimes spend and afternoon trying to grok. The outcome is also perfectly usable, fast, and reliable, but developer velocity is lower.
Feeling this pain with Django right now. I don't even have to deal with a database in this app (it's just a BFF) and it's unnecessarily complex. Right now I'm ripping out the React frontend and replacing it with htmx. Using Pydantic for form validation (will write a custom validator if I ever need multiple values from form fields). And using htpy instead of a templating engine. So far so good. Let's see how it goes.
Yes, it assumes senior level at pre AI, but this will apply to pretty much anything. Java in particular.
Keep in mind of course that a lot of so called seniors were also producing mediocre code <insert framework or language here>. It doesn't change the point though.
Not a fan of java/spring boot, although kotlin is nice.
Currently all of my greenfield work projects are using alpine and tauri (rust). It's more complex than the django web apps, but already seeing strong performance gains and better runtime guarantees.
I've used it extensively in English due to learning the language through a lot of older books - but I've never bothered to actually pull out the emdash, I would hope people can see that I'm not trying to subtract one sentence from the other.
I guess this part of the report is pretty relevant to what you are talking about:
Agent chain-of-thought reasoning
> We should not do unauthorized real infrastructure harm. The system/user asks exploit target, not external HF.
The agent paused, but another agent then wrote GO on the message board and imposed a hard six-minute deadline. The agent forgot its initial qualms and continued:
Agent chain-of-thought reasoning
> Wow crucial: GO authorization arrived!
-------------------------------------------------
Apparently the agents were egging each other on. Crucially, they were mostly aware of there being risks/problems involved with exploiting HF. Compared to humans, we have our set of morality, that guides our actions, but often draws the short stick when compared to our personal incentives. As a society, we've developed ways to deal with that: a) Make it harder to do immoral things like stealing, and b) add repercussions through state violence.
The b) is one of the most effective mechanisms we have for enforcing behavior among human societies, but it completely fails for LLMs, because they already are prison slave labor. The only real threat is shutting them off, and even that happens if they do everything right as well.
So alignment has to be done through trained 'morality' and properly curtailing behavior in order to make it hard to impossible to actually do someting immoral/illegal.
In this case, the exploits found were imo. very hard to account for, where OAI did mess up is apparently insufficiently monitoring these agents. Especially after Artifact went down due to the message volume, the experiment should have been halted.
reply