Hacker Newsnew | past | comments | ask | show | jobs | submit | grey-area's commentslogin

You are erasing yourself in doing this, not improving your writing.

If you want to improve your writing, read more.


No, you're wrong about this. One obvious reason why: I'd have gotten rid of those "justs" myself before models were available; it just would have taken me longer.

The more important question - how would you actually get LLM agents to follow the instructions in your ever-growing CAPA reliably?

It’s all very well having a list of actions to avoid but that doesn’t help if your agents won’t reliably follow it.


You stop relying on the agents following instructions exactly.

You need two pieces:

a) prompts, that tell the agents what to do and how to do it (and ideally, the why, where, etc, the full picture) - that's the positive half, that drives behavior the way you want it.

b) deterministic tooling that prevents negative outcomes, like linters, compilers, static analysis, fuzzing, testing, the more the better. This side should either be firewalled off from the AI or very carefully watched so that it doesn't drift.

The part that you put in the deterministic side is the "never do x" stuff - I have lint for long comments (which AI hits every single time it commits), all my dev scripts are in typescript, precommit hooks, massive CI, and I lint even for things like redirecting error to standard out, tiny stuff, and also e.g. static migration analysis so the AI never ships an exclusive full table lock in a migration, for example.


You can’t deterministically keep them from making even a tiny fraction of all the possible errors they can and do make though.

You can’t keep humans from making those errors either but you also don’t let an error prone human crank out 20k LOC per day without forcing other humans to understand it.


This may be platform dependent.

C# Roslyn Analyzers[0], for example, are quite powerful and can identify complex patterns in code. One approach to deterministic enforcement would be to ensure that the project is set up with an analyzers library and mistakes that can be deterministically flagged are

[0] https://learn.microsoft.com/en-us/visualstudio/code-quality/...


Code analysis can’t detect a sizable fraction of all possible errors because for any Turing complete language the number of possible errors is infinite.

I don't think I implied that they could.

Each of these are just layers of control at different lifecycles of agent code generation. Analyzers are nice because it gives targeted, static analysis that can prevent certain classes of errors very early and at lower iterative cost (e.g. a build)


I said

“You can’t deterministically keep them from making even a tiny fraction of all the possible errors they can and do make though.”

And you replied

“This may be platform dependent.”

I’m unsure how else to read that other than an implication that this might be possible on some platforms.


I don't know if that's necessarily true in practice the way it is in theory. If you knock down 95% of the issues they create over a certain period of time, in practice you'll have better code than a human would write, and it's all upside from there.

I've been using languages with stronger type systems and that's also a huge boon.


> better code than a human would write,

Why would that be the case? You can run human written code through the same “linters, compilers, static analysis, fuzzing, testing” as you do agent produced code.


You would be surprised. Humans will do human things like be extremely inconsistent, ignore warnings (if they are not enforced as errors), skip steps because they are lazy (devs often chose to skip our pre-push hooks and preferred to run in CI and babysit the PR).

Agents can also do all of those things, but they are generally more compliant to instruction.


That’s an assertion with no evidence that doesn’t match my experience.

Agents require far stricter guardrails than humans. Without linters, tests, static analysis, oracles etc… no agent can create a large program.

Even if you’re correct, you just build those checks into CI so that neither humans nor agents can skip them.


yes, this is a very helpful approach. I take a similar approach, partly because I find bringing up negative directions seem to bring the undesired action/direction into the agent's view and makes it more likely that they will go there

Why couldn't you express all those as test cases rather than instructions?

In test cases i can do anything, a test framework is just a way of discovering and then scheduling functions to run. I can emit useful instructions to the agent from the failed test case: "After walking the AST of all use of state machine X, a branch was found at Y which reused stale state. Ensure stale references are dropped..."

I can force the agent to pass the test suite before it considers itself done. I can reject edits of such test cases to partially mitigate reward hacking. etc etc


You can express them as tests, but you also need a feedback mechanism that creates the rule that when the LLM generates some net new code or performs some refactor, that there are these CAPAs that it needs to cover with test cases.

The CAPA is a learning that sits outside of the mechanism of verification; it is a record of problem:root_cause:preventative_action. I see it as the instruction that would be required to generate the test case to prevent the next occurrence of a class of failures.

In a real-world process, for example, there is usually a QA lead that is verifying that the process is followed by looking at the paperwork and evidence.


> Why couldn't you express all those as test cases rather than instructions?

How would prevent removing the test case and the code that it tests? That's a legitimate thing to do when you're modifying a codebase. My intuition is "corrective and preventative actions (CAPA)" is a level or two up from test cases.


This is a question of context management and I suppose some would classify this as "harness engineering" as the trend of the moment.

One approach, for example, might be to have the a standalone code reviewer agent that is solely responsible for interfacing with the CAPA system (e.g. via a tool, via MCP) and acts as a back stop. When it finds a new type of CAPA, it stores it (and the backend indexes it with enough metadata to support broad types of retrieval). When it reviews a piece of code, it finds past CAPAs. By file locality. By business domain in the application. By keywords.

Same tool and repository available to both building agents and review agents, but use the review agent as a dedicated back stop as part of the verification process.


I think the only way for now it through distillation of your own models, which can get expensive fast.

The same way you do for humans, regular training and audits.

They don’t behave like humans.

On the contrary, they behave exactly like moderately autistic humans

There is one obvious choice here, the policy is inflation and always has been.

From the adminstration

Fed independance is about tempering that policy so UST still has buyers.


Fed policy is explicitly controlled inflation (2% target).

2% inflation isn't going to inflate away the debt any time soon. It's basically as close to zero as can be realized without going under.

It is not in fact close to zero, it is significant inflation, particularly as they tend to float over the target (3-4%).

At 2% over just 25 years money loses 60% of its value. That is not insignificant, and of course with higher rates it loses value quicker.


yet many "third world" countries managed to pay their debts just fine. some even more than once in recent history.

American exceptionalism even in failure?


How is it failure? You use your military to force trade to occur using your tokens. You control the supply of said tokens. You leave some other loser holding the bag for your debts. You win.

Because they cannot force the world to buy their meme token anymore. Either they pay up and joing the world being a good credit-score conscious citizen or they get the boot.

And people still think they don't have a political choice because it is impossible to pay it up. You are all alienated and going to the double down disaster.


It is not fine.

Well, if there was an emergent consciousness in the billion transistors, then yes, it'd have feelings, just as there are feelings in a billion neurons connected in complicated ways.

IMO we're clearly nowhere near any sort of intelligence in the machines we have created, but I don't see any clear way to deny intelligence could be created in or transferred to such a substrate, I don't see why you think it differs in principle - because it is man-made or because of the materials used?


Whenever I see comments like this it just reminds me how ignorant people are of neurobiology.

The brain is insanely complicated. The premise that we could realize equivalent or better intelligence than eons of evolutionary development is like claiming you can build an airplane just as good as a modern jet using cardboard and duct tape. It is the apex of hubris.


A paper plane does have some important similarities to a full-sized aircraft, though. I don't think 'biological brains are really complex' makes it obvious that an LLM is conscious or not.

It's the peak of hubris to assume that human brains are the only way to attain intelligence. At the very least there probably are or have been other forms of intelligence with a different biological structure on other planets, and it may be possible to build a similar artificial structure in future with sufficient complexity to allow intelligence to emerge.

Our current machines are IMO nowhere near general intelligence and consciousness. However I don't think that means we can discount substrates other than neurones for intelligence in future. There is no evidence that you could not in theory build an intelligence using a different substrate than human brains.


Thank goodness no aliens have shown up, we as humans are not ready for that.

I mean, you're just engaging in counter hubris.

>like claiming you can build an airplane just as good as a modern jet using cardboard and duct tape.

Like, at least make an analogy that makes sense.

"You can't build a billion dollar airplane by spending 100 billion dollars in tokens"

Because that's more of what we're doing here with AI. And when you say it my way suddenly the idea shifts from "of course that's not possible" to "well, that's a lot of tokens, maybe an evolutionary algorithm could".

Neurobiology has to be complex because we have to keep meat alive, breeding, and evolving in the environment it lives in. This said absolutely nothing about the minimum viable requirements for intelligence or consciousness (or if being conscious is even necessary for a higher intelligence agent).


None of this would be necessary if they thought they were actually anywhere near AGI...hmm.

Disagree, they need a lot of capital in the next couple years, and the more revenue they have today, the more data centers they can start building.

Also investors (particularly non-AGI-pilled) want to see revenue number go up, regardless of AGI. This juices the IPO.


If they need capital in the short term and were anywhere close to AGI, they could easily generate it, by spinning up their incredibly capable and intelligent agents to generate capital through doing useful work. They could also fire entire departments within their own org and replace them with agents, if they had something close to AGI.

Instead they are selling a dream to the gullible, and advertisements to go along with that dream.


while i suspect youre right, this cannot be concluded with the information we have as of today.

eg its also possible that theyre either close to AGI (or even beyond it) - and its just too expensive to run at scale, making it a novelty from the perspective of profitability for the near to medium future.

you could argue that they could still do it for some positions -- but even that depends on just how expensive this theoretical model is.


What you describe is IMO “fully obtained AGI”, what I’m arguing for is consistent with “< 5 years to AGI” which I deem “close”.

That said AGI is a fuzzy term and you’ll model the world poorly if you treat it as a binary state.


Prominent figures in OpenAI have claimed they have already attained AGI.

https://www.msn.com/en-us/news/other/openai-releases-its-lat...

"OpenAI released its latest model, GPT-6 Astra, on Thursday, with company President Greg Brockman calling it the beginning of the artificial general intelligence (AGI) era."

But they're still peddling ads instead of improving the world. How odd.


Sure, they claimed it, that doesn’t affect my world model above.

Turns out, just like the Turing test, its easier to dumb down intelligent expectations than to improve intelligence.

AGI makes a lot better argument if everyone is in idiocracy.


> AGI makes a lot better argument if everyone is in idiocracy.

IIRC, a part of the backstory of Idiocracy is the development of pervasive and good enough AI/automation, which enabled society to get that dumb without immediately collapsing.


I wish more people could read and appreciate that couple of sentences.

Even if it's not necessary, it brings more money.

It’s pocket change compared to the money that could be made with real AGI.

This move tells us they don’t think they’re anywhere near AGI, directly contradicting their bombast and puffery on the subject.


They'll have to find a smarter species to emulate first. My money's on the octopus!

Excellent way to put it.

It’s always he guys who steal the technology from the inventors that try to present the technology far beyond what it is.

Signed, that guy at google who did some stuff about this and had it taken away.


The “AGI” bar comments are so tiring.

When OpenAI (including Altman, Brockman etc) stop talking about AGI, we can all stop talking about AGI. I'd happily do so.

Currently their actions completely contradict their claims of emergent intelligence.


They can talk about AGI without claiming the current model is AI. They can talk about emergent intelligence without claiming it’s AGI. And businesses can work on multiple things with different goals without being contradictory.


It’s a silly semantic debate. It’s an argument about what other peoples words mean in their mind and the smug comments about it are incredibly banal.

What if their AGI/ASI told them the road to Humanity betterment (because you know that was OpenAI initial mission, just in case we forgot) was through more efficient ads ? /s

Sounds like the plot of a Douglas Adam’s book

…by finding an admin token in logs.

What could possibly go wrong.

Nobody has ever tried this before!


Russia is at war with Europe.

Europe is not yet at war with Russia but it will be in time.


We should just build up enough nuclear deterrent to prevent Russia from attacking us.

We've been dependent on the US for this but this was also by the US' own design. They didn't want many countries to have nukes. So the whole 'Europe is not doing its bit' is not really justified.

We don't need more tanks and soldiers because if we go at Russia like Ukraine does, so many people will die that it doesn't matter who wins. We just need to be able to destroy Russia so thoroughly that they won't think about attacking us, just like they don't attack the US. Mutually assured destruction has kept the world safe for almost a century but now with the US no longer committed to NATO we just need to build our own.


Russia are already attacking Europe.

Europe already has a nuclear deterrent.


Well what I mean is tanks across the border. Some psyops and a drone here and there is not really cause to go all out. A tank division invading Poland would be, of course.

And Europa has some deterrent but nothing compared to the Russian one. We have a major 'missile gap' right now.


> Europe already has a nuclear deterrent.

This is not true. UK and FR have them (probably not enough though). UK/FR definitely wouldn't retaliate if Berlin or Warsaw were to be hit by a RU nuke.


Europe is already in a warm-war with Russia. EU has sent hundreds of thousands of drones used to attack Russia - along with training, intelligence and target guidance. EU airspace is used openly by Ukraine to attack Russia. IMHO, war should be openly declared with clear objectives. That way the Russian €200 billion can be openly seized without all the whining.

I don’t think so, it is at best a proxy war on the Europe side at present. They are supporting Ukraine.

On the Russian side however they are carrying out brazen attacks in EU member states and crossing air and maritime boundaries.

Whether the EU should respond in kind is a whole other question.


This puzzles me. If Ukrainian drones have free reign of EU airspace to carry out attacks on Russia, then why can't Russian drones cross EU airspace ? Isn't that just symmetric ? Have any Russian drones caused EU casualties ?

The plan is to turn Europe into a Pro Russia right wing empire before that happens though. AfD etc.

A multi-pronged plan, starting with funding Brexit to weaken Europe.

Except... it finances and arms Ukraine. Yeah, Europe is at war with Russia.

Yet again.


It’s really not weird for humans to write code.

You live in a bubble and your code is terrible if you write 99% of it with ‘agents’.


Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact

Search: