[boh]

I think the big secret is that AI is just software. In the same way that a financial firm doesn't all of sudden make a bunch of money because Microsoft shipped an update to Excel, AI is inert without intention. If there's any major successes in AI output it's because a person got it to do that. Claude Code is great, but it will also wipe out a database even though it's instructed not to (I can confirm from experience). The idea that there's some secret innovation that will come out any minute doesn't change the fact that it's software that requires human interaction to work.

[codingdave]

Yes, and it has been said since day one of LLMs that all we need to do is keep things that way - no action without human intervention. Just like it was said that you should never grant AI direct access to change your production systems. But the stories of people who have done exactly that and had their systems damaged and deleted show that people aren't trying to even keep such basic safety nets in place.

AI is getting strong enough that if people give some general direction as well as access to production systems of any kind, things can go badly. It is not true that all implementations of agentic AI requires human intervention for all action.

[Terr_]

My cynical rule of thumb: By default we should imagine LLMs like javascript logic offloaded into a stranger's web-browser.

The risks are similar: No prompts/data that go in can reliably be kept secret; A sufficiently-motivated stranger can have it send back completely arbitrary results; Some of those results may trigger very bad things depending on how you use or even just display them on your own end.

P.S. This conceptual shortcut doesn't quite capture the dangers of poison data, which could sabotage all instances even when they happen to be hosted by honorable strangers.

[jrflowers]

If you had made a tool that gave gpt-3 the ability to run arbitrary commands on your production systems you could have seen things go badly.

[Lalabadie]

Good news! Today's SOTA models can also make things go badly.

[jrflowers]

Yep. I don’t see how that metric indicates how… strong(?) a language model is.

[stuaxo]

Eh, these same people will attach openclaw to production systems soon and destroy their own companies.

[CamperBob2]

The problem is, out of ten companies who take this approach, nine will indeed destroy themselves and one will end up with a trillion-dollar market cap. It will outcompete hundreds of companies who stuck with more conservative approaches. Everybody will want to emulate company #10, because "it obviously works."

I don't see any stabilizing influences on the horizon, given how much cash is sloshing around in the economy looking for a place to land. Things are going to get weird, stupid, and chaotic, not necessarily in that order.

[flats]

One does not even need OpenClaw to achieve this outcome: https://x.com/lifeof_jer/status/2048103471019434248

[ffsm8]

Yeeeehaaaaa, the vibes shall never end!

On a more serious note, they were mostly f*cked by their paas provider imo. Claude will always do dumb shit. Especially if you tell it to not do something... By doing so you generally increase the likelihood of it doing it.

It's even obvious why if you think about it, the pattern of "you had one job, but you failed" or "only this can't happen, it happened!" And all it's other forms is all over literature, online content etc.

But their PaaS provider not scoping permissions properly is the root cause, all things considered. While Claude did cause this issue there, something else would've happened eventually otherwise.

[flats]

I absolutely agree with you.

Also, some folks seem to be forgetting the virtues of boring, time-tested platforms & technologies in their rush to embrace the new & shiny & vibe-***ed. & also forgetting to thoroughly read documentation. It’s not terribly surprising to me that an “AI-first” infrastructure company might make these sorts of questionable design decisions.

[AndrewKemendo]

Sounds like a pretty efficient self correcting mechanism

I’m not sure what the problem is there

[tikkabhuna]

The problem is that destruction isn't contained to the company. If an AI agent exposes all company data and that includes PII or health information, that could have an impact on a large number of people.

[AndrewKemendo]

PII breaches have been pretty consistently a problem for the last several decades, predating modern LLMs.

So that is a structural problem with their data and security management and operations, totally independent of the architecture for doing large scale token inference.

[ben_w]

Normalisation of deviance is the problem: https://en.wikipedia.org/wiki/Normalization_of_deviance

Remember that these models are getting better; this means they get trusted with increasingly more important things by the time an error explodes in someone's face.

It would be very bad if the thing which explodes is something you value which was handed off to an AI by someone who incorrectly thought it safe.

AI companies which don't openly report that their AI can make mistakes are being dishonest, and that dishonesty would make this normalization of deviance even more prevelant than it already is.

[AndrewKemendo]

That’s not a technical/AI problem in any sense, that’s a social problem in organizing and coordinating control structures

Further, it’s only a problem to the extent that the downsides or risks are not accounted for which again… is a social problem not a technological problem

This isn’t a problem for organizations that have well aligned incentives across their workflows

A well organized company that has solid incentives is not going to diminish their own capacity by prematurely deploying a technology that is not capable of actually improving

The issue is that 99% of the organizations that people deal with have entirely orthogonal incentives to them. They are then attributing the pain in dealing with that organization to the technology rather than the misaligned incentives

[ben_w]

> That’s not a technical/AI problem in any sense, that’s a social problem in organizing and coordinating control structures

As @TeMPOraL here likes to point out, it can be genuinely fruitful to anthropomorphise AI. I only agree with partially, that this is true for *some* of the failure modes.

> A well organized company that has solid incentives is not going to diminish their own capacity by prematurely deploying a technology that is not capable of actually improving

Sure, but society as a whole doesn't have the right solid incentives to make sure that companies have the right solid incentives to do this. We can tell this quite easily by all the stupid things that get done.

> The issue is that 99% of the organizations that people deal with have entirely orthogonal incentives to them.

This is also fundamentally the AI alignment problem, that all AI are trained on some fitness function which is a proxy for what the trainer wanted, which is a proxy for what incentives their boss gave them, which is a proxy that repeats up to the owners in a capitalist society, which is a proxy for economic growth, which is a proxy for votes in a democracy, which is a proxy for good in a democracy.

[dataviz1000]

LLM models are a distribution. Unlike a python script or turning machine, a LLM model is capable of generating any series of tokens. Developers need stop reasoning about LLM agents as deterministic and to start to think about agents in terms of Monte Carlo and Las Vegas algorithms. It isn't enough to have an agents, it also requires a cheap verifier.

If I was a Ph.D. student today, I'd probably do a thesis on cheap verifiers for LLM agents. Since LLM agents are not reliable and therefore not very useful without it, that is a trillion dollar problem.

Once a developer groks that concept, the agents stop being scary and the potential is large.

[add-sub-mul-div]

If you told a programmer 30 years ago that someday we'd switch from a deterministic to nondeterministic paradigm for programming computers, they'd ask if we'd put lead back in the drinking water.

[munk-a]

We'd just explain that management told us we had to and then they'd understand.

[dg247]

Been doing this 30 years now. I am asking that question. Everyone talks around it.

[52-6F-62]

You aren't alone.

Not even a few years ago if you introduced a component to a system that would result in non-deterministic output... Hell, a single function... You would be named and shamed for it because it went against every principle you should be learning as a novice writer of software.

I have used the LLM tools, and I see the real-world potential for these things. But how it's all being sold and applied now: it's upside down.

[reducesuffering]

Right? I get a kick out of programming used to being:

put this exact value inside this exact register at the right concurrent time and all the tedious exactness that C required

into now:

"pretty please can you not do that and fix the bug somewhere a different way"

[georgemcbay]

> they'd ask if we'd put lead back in the drinking water.

With Lee Zeldin heading the EPA is anyone sure we won't?

[goatlover]

Replace fluoride with lead in the water. Blocks out all the negative effects from wind turbines. /s

[com2kid]

It has always been non-deterministic but we relied on low level engineers who knew the dark magicks to keep the horrors at bay.

Bit flips in memory are super common. Even CPUs sometimes output the wrong answer for calculations because of random chance. Network errors are common, at scale you'll see data corruption across a LAN often enough that you'll quickly implement application level retries because somehow the network level stuff still lets errors through.

Some memory chips are slightly out of timing spec. This manifests itself as random crashes, maybe one every few weeks. You need really damn good telemetry to even figure out what is going on.

Compilers do indeed have bugs. Native developers working in old hairy code bases will confirm, often with stories of weeks spent debugging what the hell was going on before someone figured out the compiler was outputting incorrect code.

It is just that the randomness has been so rare, or the effects so minor, that it has all been, mostly, an inconvenience. It worries people working in aviation or medical equipment, but otherwise people accept the need for an occasional reboot or they don't worry about a few pixels in a rendered frame being the wrong color.

LLMs are uncertainty amplifiers. Accept a lot of randomness and in return you get a tool that was pure sci-fi bullshit 10 years ago. Hell when reading science fiction now days I am literally going "well we have that now, and that, oh yeah we got that working, and I think I just saw a paper on that last week."

[greysphere]

With the old way of doing things you could spend energy to reduce errors, and balance that against the entropy of you environment/new features/whatever at a rate appropriate for your problem.

It's not obvious if that's the case with llm based development. Of course you could 'use llms until things get crazy then stop' but that doesn't seem part of the zeitgeist.

[com2kid]

> It's not obvious if that's the case with llm based development. Of course you could 'use llms until things get crazy then stop' but that doesn't seem part of the zeitgeist.

Harnesses are coming online now that are designed to reduce failure rates and improve code quality. Systems that designate sub-agents that handle specific tasks, that put quality gates in place, that enforce code quality checks.

One system I saw (sadly not open source yet) spends ~70% of tokens on review and quality. I'll admit the current business model of Anthropic/OpenAI would be very unfriendly to that way of working. There is going to be some conflict popping up there. Maybe open weight models will save us, maybe not.

If Moore's Law had iterated once or twice more we wouldn't be having this conversation. We'd all be running open weight models on our 64GB+ VRAM video cards at home and most of these discussions would be moot. AI company valuations would be a fraction of what they are.

[danaris]

> It has always been non-deterministic but we relied on low level engineers who knew the dark magicks to keep the horrors at bay.

This is a disingenuous comparison.

First of all, what you're talking about is nondeterminism at the hardware level, subverting the software, which is, on an ideal/theoretical computer, fully deterministic (except in ways that we specifically tell it not to be, through the use of PRNGs or real entropy sources).

Second of all, the frequency with which traditional programs are nondeterministic in this manner is multiple orders of magnitude less than the frequency of nondeterminism in LLMs. (Frankly, I'd put that latter number at 1.)

This is part of a class of bullshit and weaselly replies that I've seen attempting to defend LLMs over the years, where the LLMs' fundamental characteristics are downplayed because whatever they're being compared to occasionally exhibits some similar behavior—regardless of the fact that it's less frequent, more predictable, and more easily mitigated.

[com2kid]

> First of all, what you're talking about is nondeterminism at the hardware level, subverting the software, which is, on an ideal/theoretical computer, fully deterministic (except in ways that we specifically tell it not to be, through the use of PRNGs or real entropy sources).

Malloc and free were never deterministic outside of the simplest systems.

The second we accepted OS preemption we gave up deterministic performance.

Good teams freeze their build tools at a specific version because even minor revs of compilers can change behavior.

I've used way too many schema generator tools that I'd describe as "wishfully deterministic".

Heuristics have been used for years in computer science, resulting in surprising behavior. My point is that if we ramp up the rate of WTF we are willing to tolerate, the power of the systems we can build increases drastically.

> Second of all, the frequency with which traditional programs are nondeterministic in this manner is multiple orders of magnitude less than the frequency of nondeterminism in LLMs. (Frankly, I'd put that latter number at 1.)

Building a RAG lookup system that takes in questions from the user, looks up answers in a doc, and returns results, can be built with reliability damn near approaching 99.99%.

I have seen code generation harnesses that also dramatically reduce non-determinism of LLM generated code, but that will continue to be a hard problem.

My phone camera applies non-deterministic optimizations to images I take, and has done so for years now.

GPS is non-deterministic (noisy), we smooth over the issues. GPS routing is also iffy, but again we smooth over the issues.

The question is can useful products be made with a technology. You can shove enough guardrails on an LLM interface to make it useful. That much is clear. I derive massive value from LLMs and other transformer based systems literally everyday. From the modern speech transcription systems, that are damn near magic compare to what we had a few years back, to image recognition, to natural language interfaces to search over company documents.

If we completely discard coding agents, LLMs are still an insanely impactful technology.

Those guardrails add costs, and latency. For some scenarios that is fine, but for others it isn't. Chat bot support agents implemented by the lowest bidder don't have any attempt at guardrails. Better systems are better built.

I agree that current LLMs all suffer from the problem that the control messages are intermixed with data, that is a crappy problem that the industry has known is a bad pattern for literally decades (since the 70s, 80s?). It seems like an intractable flaw in the systems.

But that doesn't make the system unusable any more than the thousand other protocols suffering from the same flaw are unusable.

[dataviz1000]

The single best example is for this discussion is Superscalar out-of-order execution which can't be used in aerospace, medical devices, and industrial control systems, or you need to guarantee that code finishes within a certain time, because technically it isn't deterministic.

Neither is stochastic gradient descent which is the cause of the LLM problem. Nor is UDP, the network protocol that powers video calls, live streaming, and online gaming.

[aleph_minus_one]

> If I was a Ph.D. student today, I'd probably do a thesis on cheap verifiers for LLM agents. Since LLM agents are not reliable and therefore not very useful without it, that is a trillion dollar problem.

PhD thesis are for (ideally) setting up a new world standard in some research area (at the end, you build your PhD thesis out of the deep emotional shards of this completely destroyed life dream), and not for some personal self-discovery project of which you hope that it will turn you into the popular kid on the block.

[dataviz1000]

That is like telling students to never do a PhD thesis on superscalar out-of-order execution, stochastic gradient descent, or UDP. I'm framing it as an analogous problem. What is missing is a cheap verification process.

[aleph_minus_one]

> That is like telling students to never do a PhD thesis on superscalar out-of-order execution, stochastic gradient descent, or UDP.

No decent PhD advisor would let their PhD student base their PhD thesis on such well-known concepts: a doctoral study programme is a journey into something never-seen-before (with a very high likelihood of faling and shattering your life). Anything else is failure.

(Obvious exception: either he or the PhD student can convince the other one that there could be something really, really deep in, say, "superscalar out-of-order execution", "stochastic gradient descent" or UDP be found that generations of researchers overlooked, and which once discovered might necessitate rewriting all the standard textbooks about this topic).

[throwaway27448]

What would a verifier even look like without having all of the same problems that the chatbot itself does? Are humans themselves not the cheap verifiers?

[xdavidliu]

humans are probably the least cheap thing you can have in this context

[throwaway27448]

Yea, but they'll do the job. What else plausibly could? ...an LLM? Then you're back at unreliable computation.

[drBonkers]

Do you have any readings you recommend to start thinking in terms of non-deterministic algorithms and cheap verifiers?

[f1shy]

Neurosymbolic programming

[whatever120]

That’s not a particular reading

[mistrial9]

filters

[airstrike]

While you're at it, I'll take a pair of unicorns too if you can find them.

[cmdrk]

My observation is that the true believers really don't want to think of models as an inert pile of weights. There's some mysticism attached to imagining it's the ship's computer from Star Trek, HAL-9000 or C-3PO. A file loaded into memory and executed over is just so... _pedestrian_.

[ben_w]

Canonically, the Star Trek computers have pretty much always been just computers, not themselves sentient because the software running on them just isn't.

I'm still not sure if HAL-9000 was supposed to be conscious or just an interesting plot device with a persona as superficial as LLMs are dismissed as today.

LLMs could definitely play the part of all three of your examples, given the flaws they showed on-screen. Could even do a decent approximation of Data (though perhaps not Lore without some jailbreaking).

Still weird that even the best of them isn't really ready to be KITT.

[bellBivDinesh]

The specter of AGI helps them obfuscate this

[cyanydeez]

I think the market isn't for anyone but other businesses. We're all ants trying to understand how AI is going to eradicate the lower levels of society.

[trolleski]

Just call the errors 'consciousness' and keep selling those tokens! Let the Spineless Generation have their last bubble!

[ctoth]

> doesn't change the fact that it's software that requires human interaction to work.

Have you ever seen Claude Code launch a subagent? You've used it, right? You've seen it launch a subagent to do work? You understand that that is, in fact, Claude Code running itself, right?

[simonw]

I don't think subagents are representative of anything particularly interesting on the "agents can run themselves" front.

They're tool calls. Claude Code provides a tool that lets the model say effectively:

  run_in_subagent("Figure out where JWTs are created and report back")

The current frontier models are all capable of "prompting themselves" in this way, but it's really just a parlor trick to help avoid burning more tokens in the top context window.

It's a really useful parlor trick, but I don't think it tells us anything profound.

[ctoth]

The mechanism being simple is the interesting part. If one large complex goal can be split into subgoals and the subgoals completed without you, then you need a lot fewer humans to do a lot more work.

The OP says AI requires human interaction to work. This simply isn't true. You know yourself that as agents get more reliable you can delegate more to them, including having them launch more subagents, thereby getting more work done, with fewer and fewer humans. The unlock is the Task tool, but the power comes from the smarter and smarter models actually being able to delegate hierarchical tasks well!

[otabdeveloper4]

You misunderstand.

The only reason to launch subagents is to avoid poisoning the LLM's already small context window with unrelated tokens.

It doesn't make the LLM smarter or more capable.

[suttontom]

Wtf? A sub-agent is a tool you give an agent and say "If you need to analyze logs delegate to the logs_viewer agent" so that the context window doesn't fill up with hundreds of thousands of tokens unnecessarily. In what universe do you live in where that mechanism somehow means you need fewer humans?

Do you think this means "Build a car" can be accomplished just because an LLM can send a prompt to another LLM who reports back a response?

[fnoef]

My Linux server runs a cron job, that can spin off a thread and even use other ~apps~ tools. Did I invent AGI?

[ctoth]

Does your Linux server decide what processes it should launch at what time with a theory of what will happen next in order to complete a goal you specified in natural language? If so yes, I reckon you sure have!

[balls187]

Claude does not have a "theory" of anything, and I'd argue applying that mental model to LLM+Tools is a major reason why Claude can delete a production database.

[Jtarii]

Well, humans also routinely accidentely delete production databases. I think at this point arguing that LLMs are just clueless automatons that have no idea what they are doing is a losing battle.

[timacles]

They’re not clueless they just don’t have a memory and they don’t have judgement.

They create the illusion of being able to make decisions but they are always just following a simple template.They do not consider nuance, they cannot judge between two difficult options in a real sense.

Which is why they can delete prod databases and why they cannot do expert level work

[liquid_thyme]

I like to think of LLMs as idiot savants. Exceptional at certain tasks, but might also eat the table cloth if you stop paying attention at the wrong time.

With humans, you can kind of interview/select for a more normalized distribution of outcomes, with outliers being less probable, but not impossible.

[freejazz]

When you're applying reasoning like this, sure, why not? What difference would it make?

[californical]

I mean maybe it’s a losing battle today, but it is correct. So in a few years when the dust settles, we’ll probably all be using LLMs as clueless automatons that still do useful work as tools

[parliament32]

So... systemd is AGI now?

[recursive]

Maybe. But probably not. It doesn't matter if it's AGI though. If those other apps and tools do simple things that are predictable, then we can be pretty sure what will happen. If those tools can modify their own configuration and create new cron jobs, it becomes much harder to say anything about what will happen.

[munk-a]

Most of us work on software that can modify its own configuration and create new jobs. I, too, have worked in ansible and terraform.

The key break here is the lack of predictability and I think it's important that we don't get too starry eyed and accept that that might be a weakness - not a strength.

[ahoka]

Well do you make 100 billion bucks with it? If no, then not AGI.

[xboxnolifes]

My claude has never yet launched itself from my terminal, gave itself a prompt, and then got to work. It has only ever spawned a sub-agent after I had given it a prompt. It was inert until a human got involved.

If that is software running itself, then an if statement that spawns a process conditionally is running itself.

[islandfox100]

Substance aside, I feel this comment is combative enough to be considered unhelpful. Patronizing and talking down to others convinces no one and only serves as a temporary source of emotional catharsis and a less temporary source of reputational damage.

[boh]

You're using it and if someone else was using it the output would be different. The point is really that simple.

[echelon]

All AI requires steering as the results begin to decohere and self-enshittify over time.

AI in the hands of an expert operator is an exoskeleton. AI left alone is a stooge.

Nobody has built an all-AI operator capable of self-direction and choices superior to a human expert. When that happens, you'd better have your debts paid and bunker stocked.

We haven't seen any signs of this yet. I'm totally open to the idea of that happening in the short term (within 5 years), but I'm pessimistic it'll happen so quickly. It seems as though there are major missing pieces of the puzzle.

For now, AI is an exoskeleton. If you don't know how to pilot it, or if you turn the autopilot on and leave it alone, you're creating a mess.

This is still an AI maximalist perspective. One expert with AI tools can outperform multiple experts without AI assistance. It's just got a much longer time horizon on us being wholly replaced.

[DeathArrow]

A one liner shell script can run itself.

[recursive]

One liner shell scripts can be analyzed. Some of them can be determined to not delete the production database. The others will not be executed.