Rich Heimann (@rheimann)

The fact that they threw it up on GitHub suggests (at least to me) that 1. they are excited about the results and 2. the work lacks the kind of rigor required for peer review; maybe even preprint. The use cases do seem arbitrary, shallow, and cherry-picked. Nevertheless, the idea that you can directly program llms directly with language using another llm is a paradigm change.

Rikiya Takehi @rikiyatakehi Nov 2

We asked GPT-6 Astra to edit Qwen3.5-0.8B weights to improve math.
It directly changed only 0.016% of weights, fixing 0.8B's reasoning loops & improving math up to 6x and even instruction following 3x, without gradients or training.

We call this Weight-Space Programming: frontier LLMs can now natively read/write neural network weights! 🧵
Blog: rikiyat.github.io/blog/weight-sp…

View quoted post on X

Rich Heimann @rheimann Nov 2

This is an interesting argument by @Kpaxs: "the output is not checkable at a lower cost than producing it."

Today, producing an answer is much cheaper than establishing that it is good enough to use. Consequently, verification may be the constraint on automation, not generation, which may function as a kind of verification or assurance debt.

But verification costs are not fixed.

I have seen a lot of activity recently with people working with SMT solvers that can check whether formalized claims satisfy explicit constraints. Ontologies to make the concepts and relationships those checks depend on explicit. Environments, harnesses and evals make testing behavior repeatable. Even models like Jev approach the problem through cheaper, structured, probabilistic judgments.

These all offer different kinds of assurance. While none makes verification free or eliminates the need to validate the checks themselves, they do shift the economics. I expect those economics to rebalance, although unevenly across tasks. But verification doesn’t need to become cheaper than generation. The combined cost of producing, checking and correcting the result needs to beat the original workflow at an acceptable level of risk. Organizations cannot postpone the infrastructure that makes this verification manageable.

Kpaxs Verified account @Kpaxs Nov 1

Buckminster Fuller, who was wrong about a lot of things in a way that was usually more interesting than other people being right, noticed something in the 1930s that still holds true.

He noticed that technology does not primarily get bigger. It gets thinner.

The transatlantic cable got lighter per unit of message. The building got lighter per unit of enclosed volume. The engine got lighter per unit of horsepower. Each generation of a technology tends to accomplish the same function with less stuff, and then with less stuff again, until at the limit the stuff approaches zero and only the function remains.

He called this ephemeralization : doing more and more with less and less until eventually you can do everything with nothing.

And here were it gets interesting. What got eaten, in the first wave, was objects: my door key became a fingerprint.

What's happening today is the second wave: the AI paradigm shift. It’s our thought patterns, our decisions, our generative processes, our creativity, and our communication and coordination with each other that is being ephemerized.

And I think judgment abstraction has two consequences.

First, the output is not checkable at a lower cost than producing it. To verify a summary you must read the document. To verify a strategic recommendation you must do the strategy. The thing that made calculators safe (verification being orders of magnitude cheaper than generation) simply does not hold for the class of work being absorbed now. Which means verification won't happen because the economics forbid it. A check that costs as much as the work is not a check. It's a second attempt.

Second, the stairs are being removed as a consequence of using the lift. The way you stay able to judge a draft is by having written drafts. The capability is maintained by its own exercise and by nothing else. So the fallback and the thing it's a fallback for are the same muscle, and the system that replaces the muscle also stops the training.

This is Lisanne Bainbridge's irony of automation, which she wrote down in 1983 and which the aviation industry spent the next forty years rediscovering at great expense: automate the easy 99% and you have designed a job where a human with no recent practice must handle the hardest 1% with no warm-up, at the worst possible moment, on the basis of a mental model the automation has been maintaining on their behalf.

Airlines eventually responded by mandating hand-flying. No knowledge-work organization on earth currently mandates hand-thinking.
View quoted post on X

Rich Heimann @rheimann Oct 26

Who is the brian scalabrine of ai researchers?

Rich Heimann @rheimann Oct 24

Hubert Dreyfus warned against treating every advance in AI as another point on a single, inevitable trajectory toward intelligence. Following Yehoshua Bar-Hillel, he called this the first-step fallacy. Early success is taken to establish the entire ambition. But improvement on one problem does not establish that the same methods will solve every other problem.

The shape of that warning helped Gary Marcus frame deep learning (DL) in 2012. He describing DL as “at best, only a small step” toward intelligent machines. Subsequent developments have shown AI to have converged around shared deep learning. Marcus was wrong.

There is an inverse error we might call the no-step fallacy. Skeptics treating the absence of a final achievement as evidence that intermediate achievements do not count. If every success can be dismissed because it leaves something unsolved, skepticism becomes impossible to disconfirm. An imagined endpoint cannot determine whether any observed accomplishment deserves recognition. While the optimist assumes the destination validates every step; the skeptic assumes that failure to reach the destination invalidates every step. Both mistake their conviction for evidence and good judgement.

Consequently, Marcus’s judgment should be evaluated against the subsequent progress of deep learning. Needless to say, it has aged less well than the technology he questioned. And continuing to identify limitations doesn't vindicate an earlier assessment that underestimated deep learning's potential. Every unfinished technology has limitations. A forecast can be wrong about the significance of a development even when that development remains incomplete.

The first-step fallacy remains a good reminder about inevitability; while its inverse is a strong warning against permanent dismissal. Both optimism and skepticism should be accountable to what a system can do, how reliably it does it, and how far those abilities extend beyond their original setting.

Rich Heimann @rheimann Oct 24

wild

Rich Heimann

Latest posts 5