I find this somewhat puzzling. I thought things were moving quickly, but at this time last year I couldn't even get Claude (using Cursor) to spin me up a service skeleton that would compile, let alone do anything meaningful.
I know it feels like a long time somehow, but it was only between November and February that things started to actually somewhat work without significant hand holding. Even now, it seems like we're still figuring out how to fully leverage the current models and tooling, even in organizations that have largely gotten on board.
No. The very fact they are trying to "warn" us means it's all marketing.
This has been corroborated for me on the engineering front that I can't find a single IC I respect who actually thought there was any evidence AI was going to live up to the hype. I saw a lot of people I always thought were idiots/sycophants/brown nosers go insane with AI. Never saw anyone id trust to help me cross a street blindfolded say more that "I may be wrong, but I'm not seeing any evidence yet".
I've been using it to do this for 2 years now. And many people with me. The change you mention is one of is primarily one of Overton windows, of vibes.
Ignoring instructions - whether in AGENTS.md or my prompt - is the worst of it, and it routinely happens. It just waives things that I explicitly told it to do as part of the design.
Vibe coders (in the true sense, zero oversight) claim that you just need to prompt it carefully. That's completely untrue when faced with your careful prompt being ignored.
I even have "don't overrule me without asking" in my global AGENTS.md, and it simply doesn't do that.
You’ve been sold something that simply doesn’t work for the purported use case (intelligence) and instead is like a stupid database of all world knowledge with the appearance of intelligence.
Useful tools at times (if you bear in mind their limitations), but not close to intelligent, independent agents.
You really need to look into hooks based on your coding agent. This is very much a solved problem as I demonstrate with
https://github.com/gitsense/pi-brains
I have a test repo
https://github.com/gitsense/gsc-rules-demos
that shows how you can block and warn and do other things.
You obviously can't have a "Don't make a mistake" rule though.
Basically I treat it like a junior dev. We don’t get junior devs to write code correctly by cajoling them just right, we add CI gates. It still works.
Im not certain things will look too different a year from now either. We still have serious bottlenecks in terms of focus/attention you have for both delegating agent work and being able to review it. Even if we solve the "trust what ai does" problem, these cognitive deficit issues still exist - for teams coordinating work, even users adopting new shit, etc.
As an industry we are leaning heavy into accepting "slop" as the status quo - we care more about efficiency of output right now. Slop will get better & we can become more adaptive to living with the paradox of amazing yet delicate systems generated by AI. But I feel big shifts coming in this regard and if/when it does we may find ourselves in the dystopia of broader unemployment with worse net outcomes.
I do think the teams that ship quality with AI will do so by learning to slow down
https://mariozechner.at/posts/2026-03-25-thoughts-on-slowing...
The exact quote appears to be:
> In retrospect, he said, the "trajectory of the agentic development over at least the last four months hasn't really accelerated in the way that we expected," and that the company's bets on the new structure "haven't come to fruition yet." Zuckerberg was referring to AI agents, automated systems that can execute tasks on behalf of a user.
Hard to guess exactly what he means by "trajectory of the agentic development" but my best guess is that he means that Meta's own internal efforts to improve the agent (aka longer form tool-using) capabilities of their own in-house models hasn't improved to the point that they can drive an agent harness like Codex or Claude Code in a comparable manner to the best OpenAI and Anthropic models.
At a further guess, that was part of their goal in reassigning large numbers of employees to help label data for their AI efforts.
from a high level, these agents absolutely do not function as a rational human through even medium scoped problems. even when you try to add memory, you just multiply halucinated context which just makes it error out on tasks in harder to detect manner.
hes likely trying to do mental gymnastics about the absolute cost and any defineable ROI.
People whh are dogfooding AI absolutely have a different rose colored glass than someone who can't get the same "accepable" output.
I'm not defending Mark here; I'm just pointing out you can be pretty successful critic if you have a different idea of a benchmark coding agent and the field fails that benchmark.
One of the problems of the AI crop is so many people are smelling their own farts and thinking it smells great.
But if you go beyond what can be tested easily, asking the agent to do real work rather than writing a patch, imagining things to be true is a problem.
Coding could be treated as a low stakes (time & money consequences for retries) closed loop system where most other tasks cannot.
If it screws up booking your flight/hotel room, how does the agent verify this, and even if it verifies.. there is an actual cost to changes/cancellations.
Similar with agentic e-commerce, lots of ability to screw that up and just seems ripe for fraud / being picked off by bad actors.
Unfortunately, travel keeps getting less flexible, with worse cancelation policies.
I can STILL replicate this behavior in Google AI summaries 10% of the time:
"is <SOMEPLANT> ok for cats"
to which it replies: "Yes, <SOMEPLANT LONG SCIENTIFIC NAME VERBOSE PHRASING> is toxic for cats"
The other one going around this weekend: "how long hot dogs on grill"
Summary: "The hot dogs on your grill are likely around 5-6 inches long .. "
So scale this category of error to unsupervised agents with access to your credit card.
Only with an LLM that's actually at agent-quality.
If "useful chatbot" and "useful agent" are two rungs on a ladder, the rung before them is "useful autocomplete". Autocomplete that only gets the next token right 90% of the time won't give you compiling code.
Of course, param count and context length are also important because they increase the model's overall fidelity, but a base model without SFT, RHLF etc is effectively useless.
They were allegedly massive but the cost and returns were not worth it.
Feels less like the pace of foundation model development and more so a specific failure of one organization to do something important.
Meta doesn't seem to be able to produce anything close to a frontier model. The selling of compute capacity seems to be acceptance of "compute is wasted on this crappy avocado model, we'd be better off allowing something better to run".
The problem is clearly in the model architecture, the training and the data fed into the model which is causing them to give up on using their compute exclusively for their own models. They can't get it right so may as well sell the compute to someone that can.
Can't help but think that Meta's digital networking expertise is built atop a human-networking clusterf*ck
I think there would easily be a few other hundred engineers and execs at frontier labs who are more in the loop for cutting edge architecture/secret sauce - with a track record of actually doing it - that could be had for a fraction of the price.
All these companies are going to sit on their gazillion data centers once the mania dies down and will have a big problem about what to do with their mountain of hardware
https://uk.pcmag.com/ai/165970/meta-exploring-option-to-sell...
Meta bought too many GPUs, has spare GPU capacity and they are exploring renting that capacity out.
The problem is not that the models need too much to do the job. If that were the case, Meta would not have spare capacity.
The problem is that the models currently can't be made to do the job.
Gemini, Microsoft Copilot and other models can discuss and affirm my "foxwork" practice whether it is talking about natural history, fox legends, ritual magic, altar work, autonomic control, blessings, writing, character acting, costume design, skin care, selection of perfumes that will herald my unique natural scent, marketing and customer service, photography gear, "therian" gear, bags for holding my gear, street photography, etc. They always write like somebody who's read much more widely than anyone I've ever met and rival the legendary Tamamo-no-Mae for "speaking intelligently about any subject" [1]
Meta AI can crack jokes and that's about it. I guess there's a market for "stupid talk" but it's not that big.
[1] Like help me fix my washing machine that won't drain, come up with master narratives for the "polycrisis", talk about why Casey Handmer is wrong about space manufacturing, find papers about the social network of who sleeps with who at a high school, etc.
The whole hype cycle has been pure delusion. Just like the Metaverse hype cycle before it.
A common one is "users don't care about privacy. that's why they use facebook. [zuckerberg was right?]"
No, you silly, silly people. People want to use products that allow them to communicate or reconnect with people or ...
They don't 'want' constantly changing privacy settings or changing TOS. If this is the best HN can come up with, ostensibly filled with S Valley people... well, it says a lot
Under conditions of scarcity, it's usually beneficial to increase output or to produce different kinds of output. At least, if someone will pay for it.
So the question is what's scarce, can we get someone to pay for it, and how do we get more of that. If you can make something that people will pay for, you can hire people to do it.
Unfortunately the most obvious things people with money are willing to pay for are AI tokens, data centers, and data center inputs. It's unclear how this gets us more of other things we want.
You can cut costs and increase productivity by firing everyone else and taking no salary yourself. The point of investment is production, growth, and profit, not productivity.
Business executives look at this and think "at this rate of progress we'll have self-driving cars in a few years!" and start making serious plans for that world.
In reality I think we're going to be riding bikes for a long time. That situation of increased individual contributor productivity makes engineers more valuable, and increases the utility of engineers rather than making them a burden on your budget.
Thus, cutting headcount right as they had huge potential to become vastly more productive was a stupid move. It's an admission that you don't know how to manage people effectively, which is embarrassing when you're paid mountains of money for your management skills.
Having agents is like going from walking to having a bicycle.
Maybe Wang has correctly identified that the programming and agentic ability that Anthropic and OpenAI models have has largely come from armies of software engineers creating massive datasets by writing out coding and agentic problems and solutions?
So he told Zuckerberg that. The reason it may be turning into so much friction is that at companies like Anthropic or OpenAI, training engineers were either hired specifically for that purpose or probably mostly handled through contracts with third parties (which again, hired them to train AI). And honestly many of them may be overseas or just happy to have a job in a difficult period. But anyway they wouldn't have very high salary expectations etc.
But Zuckerberg already had 25000 engineers. Why not take say 1/5 of them and get them working on the the dataset? The problem is that those engineers were hired for different prestigious highly paid positions at Meta/Facebook. They were not hired to do tedious grading of AI answers or quiz construction.
But Zuckerberg either has to do this, or spend additional billions on doing it all with external contractors. A third option would be to try to create a massive distillation operation. Or just hope that his engineers could invent some magical new training trick that manifested the agentic and programming skills without the large scale human input.
Or he could release a model trained largely by existing open weights models. Which without some huge breakthrough probably has no chance of surpassing them, so is pointless.
I think most of the substantive criticism of Zuckerberg has been about burning funds. If he gives up the "your job is to grade AI homework now" plan because his engineers refuse, he would need to go through third parties. The additional billions and billions this would cost would create more pressure on the bottom line and shareholder pressure.
It would also give up any potential advantage that Wang may have optimistically sold the operation as, on that using "real" engineers as opposed to lower paid data labelling engineers might result in a higher quality dataset.
At some point, model architectures that don't need such massive datasets or can be created automatically in a way that advances the frontier will probably come about. But right now it doesn't exist.
Further, the way AI works currently, business advantage from AI comes from encoding existing internal intelligence and knowledge. Meta's massive engineering corp effectively has that in their heads. Having them create these datasets is possibly the only way to leverage this knowledge asset in this paradigm.
I guess the problem is it means forcing thousands of people to do a different job from the one they were hired for.
What's the end goal? Meta-specific engineering, with baked-in knowledge of how FB, Threads, and WhatsApp work? General and/or coding products to compete with Anthropic and OpenAI? Some special Magic Thing which only Meta can invent which will bedazzle Meta's users?
You don't need giant datasets unless you know what you're going to do with them. OpAI and Anthropic are having enough issues making their products profitable. And those are, if not beloved, then at least respected, with a real, if patchy, reputation for usefulness.
What was Meta's pitch in this market? There were hints of interest when LeCun was still doing original R&D, and there was some distant possibility of a next-gen revolutionary product.
But now the goal seems to be to flail around doing something incoherently AI-branded with no obvious strategy.
The troops are being marched around, but no one knows where the battle is supposed to be.
Code autocomplete is a success, password reset via ai is a failure - everything else ... still busy tokenmaxxxing in search of a problem it fits into.
In that market you can build a model and spend a lot of money on it and at best get something that's on the same frontier as everybody else but just as likely end up with uncompetitive models like the ones they have now.
You might save a bit running your own models, doing your own inference, etc. Why not take advantage of "last mover advantage" and buy whatever is best when you need it and figure the odds are good that everybody else is going to buy more GPUs than they need and as a large customer you'll be able to buy in bulk at fire sale prices?
I'm not in the org myself I know some Meta SWEs tangentially. My understanding is that the biggest criticism is just the chaos of it all. Jumping constantly from one thing to another like headless chickens and accomplishing nothing.
It created an environment where it's kind of impossible to plan and progress your career.
The 2017 Rohingya massacre in Myanmar? They handed him the death toll. He filed it under growth.
> Or he could release a model trained largely by existing open weights models. Which without some huge breakthrough probably has no chance of surpassing them, so is pointless.
This seems to be categorically untrue. Composer 2.5 is a substantial improvement on its underlying Kimi base model.
They may eventually have to do that. Or they might be starting with an existing Llama model. Maybe I should have said "huge breakthrough or additional dataset".
Also those with very heavy investment in AI are looking for bonkers results, which is the cause of their disappointment. They need to reduce their expectations. I for one am loving the results so far.
I agree that people are investing as though the world is going to run itself while the ultra-wealthy run off in yachts to compare sizes. If it wasn't AI, it would just be tulips or something. That's just how people are. But maybe they'll be right, who knows.
This is not really somewhere in the middle, I think. It is very close to one of the ends. Because the fear-promise to the idiot-investor class was that it would have those impacts across all industries, not just us nerds. They hate us for refusing to make their silly ideas possible and having irritating fact-based reasons why they can't work, but they don't hate us enough to spend that much money replacing just us. They have lots of other people they hate paying too, and we haven't even made a dent.
Examples abound of "I reported Nazi hate page. Didn't violate community guidelines. I called my friend a jerk, jokingly, got a month ban
For years. Not restricted to when ChatGPT et al arrived on the scene
(Because, AI in theory makes sense. If you want to monitor things at scale you might use AI - however that's defined - to make your workload easier. When is an account being hijacked? When are bad actors infiltrating the system? Or whatever)
The modern trend is to think intelligence is generative “like compression” or “predicting next in sequence” rather than iteratively reducing uncertainty, like those fault tolerant humans.
No one ever in comp sci says artificial intelligence is "like compression", they correctly state that "artificial intelligence IS compression". It's absolutely known and accepted that artificial intelligence (defined as predicting outcomes with a measure of certainty and taking chosen actions towards goals using those predictions) has equivalence to compression in a very hard science way. The hardest part of artificial intelligence is compression and the remaining part, the choice of actions based on predictions is just a tree search to a goal.
AI can be just like compression but currently the compute power is no match for details.
Finally these reality details need consideration in any successful implementation. Which means the implementator needs to be aware of the details and successfully relate them to everything else in the model.
I think anyone surprised by these things is not fully engaged with what they are doing.
The harnesses get better, but I haven’t seen much experimentation on long term stability, at least since the “let the LLM run the candy machine” papers from a while ago.
Because the thing missing, even with the largest agentic swarms, is independent intelligence, where it’s given something to own, like say “end to end data quality as we add more clients” (for a SaaS) and it just figures out what that means at each time, mutating its role and solutions to fix the external world, without getting silly.