Rob

Engineer, working on stateful agents.

[Rob @rrobmalone]

GPT-5.6 still needs to be steered and babysat in a way Fable doesn’t, unfortunate

[Rob @rrobmalone]

@signulll The “what it can do” type reviews are almost useless now, the thing I find valuable is insight into the “character” of the model, i.e. sycophancy, resourcefulness, determination, etc. these do tend to be similar user to user in my experience

[Rob @rrobmalone]

It remains noteworthy to me how many resources the labs continue to pour into UX despite recent step changes in publicly available intelligence, never mind what they have internally.

Maybe I’m over-indexing on this, but every time I see a “cowork” type product or a chat bot release with a fresh coat of paint, it reinforces my belief that we’re still firmly in the “intelligence as a tool” paradigm

[OpenAI Developers @OpenAIDevs]

We’re bringing Codex and ChatGPT together in one desktop app.

The same powerful coding agent, now alongside ChatGPT Work, with new coding workflows, a new Chrome extension, revamped in-app browser, and faster Computer Use powered by GPT-5.6.
https://t.co/9hOtxyDEw5

[Rob @rrobmalone]

@LyudaKozlovska Much of history is people learning the same lesson over and over again - consolidation of power leads to bad outcomes

[Rob @rrobmalone]

Europe is simultaneously converging toward a draconian surveillance state while regulating out the tech companies that will enable it; truly a lose-lose situation. Where did it all go wrong?

[Lyudmyla Kozlovska 🇪🇺🇺🇦 @LyudaKozlovska]

The vote is over — and on paper, we lost. Chat Control 1.0 is reinstated until April 2028.

But look at how it passed: with just 314 votes in a 720-seat Parliament — because the EPP’s second-reading trick meant opponents needed 361 to stop it. The bar to block it was set higher than the bar to pass it.

It carried on the last day before summer recess, with barely 600 of 720 MEPs in the room. What Parliament rejected in March passed today — on an empty chamber and a procedural tilt. That should trouble anyone who cares about how democracy works, not just about privacy.

What this means in practice: US tech companies are again allowed to scan private messages without a warrant or prior suspicion — direct messages on Instagram, Discord, Snapchat, Skype and Xbox, and emails via Gmail and iCloud.

The real fight now moves to September, when Chat Control 2.0 — the permanent regulation that reaches into end-to-end encryption — returns to negotiation.

[Peter Gostev @petergostev]

Fable is a dangerous model, it is such a clear and convincing writer that it is even more tempting to agree with whatever it is saying. But every time I pay attention, I find things I disagree with and inevitably get 'you are absolutely right' back. Remember, it is not AGI yet.

[Rob @rrobmalone]

@teortaxesTex Please, can we stop watering down the definition of AGI yet? Fable consistently gets things wrong, if it’s not for you you’re self-reporting a skill issue

[Rob @rrobmalone]

@thsottiaux Code editor/file explorer is still not great (maybe not top prio but I still read my code 🙃)

[Rob @rrobmalone]

@beffjezos

[Rob @rrobmalone]

@tenobrus @benryanwriter The last real people on here are the anime pfps, what a strange twist of fate 😭

[Rob @rrobmalone]

I think this misses the incentive structure, especially for the “good enough” models. Once models start generating value internally for the labs beyond just executing someone else’s prompts, e.g. automated AI R&D or scientific discovery, the dominant incentive is to push the frontier and reinvest all output into the next version, almost by definition.

IMO we’ll know we’re in the end-game when labs put their money where their mouth is on RSI, stop spending as much time and resources on inference and distribution and start using internal models for genuine value generation rather than making the best chatbot tool/UI.

The strongest sign to me that we’re not close to RSI is how much of the labs’ marginal energy still seems to be going into inference

[Rob @rrobmalone]

@theCTO @AnthropicAI @OpenAI Maybe the real AI psychosis was the performance degradation we imagined along the way

[Rob @rrobmalone]

@petergostev True but simultaneously I find it hard to follow sometimes when it’s gets comfortable in a repo, almost seems like it’s making semantic connections and shorthand in its inner monologue/COT without giving me a legend to understand it until I ask for clarification

[Rob @rrobmalone]

@omarsar0 Everyone went silent because we’re building with Fable

[Rob @rrobmalone]

@kimmonismus The limits for subscription customers and Fable API prices are only reinforcing their image IMO

[Rob @rrobmalone]

Increasingly obvious that enterprise agents will split into two layers:

Frontier models design the workflow, define the state, and build the harness upfront

Light, cheap models execute inside that harness repeatedly

X-high effort for every task is untenable

[Hedgeye @Hedgeye]

BREAKING: Meta is spending close to $50,000/year per employee on AI tokens.

[Rob @rrobmalone]

@synthwavedd Excited for Fable 5 part 3: the reckoning

[Rob @rrobmalone]

@dee_bosa @zerohedge They’re selling it because there’s demand, what’s this have to do with compute scarcity? It is scarce, and they happen to have it. Not selling excess compute would be leaving billions on the table for both

[Rob @rrobmalone]

@AnthropicAI Fallback to Opus for… coding?

[Rob @rrobmalone]

I’ll try to crystallize how I’ve been thinking about this. I think the effectiveness RLVR post-training is constrained by three things:

1) The validation loop

RLVR works when the training loop can be designed with this shape: generate an answer, test it, reward the model if the result satisfies the verifier. This is why code and math are such natural fits - the sandbox is rich, the rules are mostly self-enforcing, the relevant state IS the context, and outcomes are cheap, so they can be directly used as reward signals. The code compiles or it doesn't, the proof step is valid or it's not, etc.

Compare this to a domain like biology. The environment is hard to bound and design, state estimation is more than half the battle, constraints and interdependence are a mess to deal with, and feedback/outcomes arrive slowly. I don’t think there is a way for RLVR to generate a strong training signal for these kinds of problems in the general case.

2) The search manifold

RLVR post training usually does not incentivize search over the entire space of possible actions. In most cases, it actually does the opposite: it constrains the search manifold so that models output the reliable path more often. This is very useful when the goal is correctness through following some known procedure, but it is fundamentally limiting when the goal is discovery.

In domains where the verifier encodes current human assumptions, this methodology will tend to asymptote toward human consensus, because the loop is rewarded for following that consensus. This is the difference between using classical mechanics correctly and overthrowing it the way Einstein did, or learning to apply Magnus Carlsen’s go-to strats and surpassing him.

3) The model architecture itself

Current LLMs can simulate action through chains of thought, tool calls, and multi-step reasoning, but the underlying policy is still mostly fixed at inference time. The model can search within the distribution it has learned, but it does not naturally become a stateful learner in the AlphaZero sense: acting, updating, and improving its policy through interaction with an environment. I'm also skeptical about "continual learning" in the traditional sense for nonlocal LLMs, the signal to noise ratio of useful training signal is so minuscule during normal use that I don't think it's useful for backprop (I could be wrong).

The current RL techniques are mostly about increasing SNR when the correct trajectory is already somewhere in the model’s learned probability space, but they do not, by themselves, move the model toward the optimum when the optimal move is not already in that space.

I do think current methods can get us to something like the new definition of AGI: an agent that can do most economically valuable tasks. If they don’t, I think it’ll be more because of diminishing returns than technical incapability. But don’t expect this trajectory to produce the Nick Bostrom-style ASI that recursively self-improves into some unrecognizable kind of intelligence. I think it is probably really good that we didn't get that first, as you argued in your original tweet.

To attempt to put words to what I think the missing pieces could be:

1) Hard, canonical state

Agents need hard, verifiable state rather than the floaty context window state we have been using. We may need to make the agent’s world-state explicit enough that learning can happen over changes in state, not just over pass/fail outcomes.

2) A general learning kernel

We also need a general learning kernel that can widen the search manifold again rather than constrain it like current approaches to truly break free of "consensus" the way AlphaZero did. This is the kind of thing that is easy to point at but hard to actually characterize within the LLM paradigm, at least for me. I suspect this is the element that Altman and Demis Hassabis are referring to when they say we are still a few breakthroughs away from true recursive self-improvement.