zby (@zby) profile picture
Joined May 2007
Photo by zby (@zby)Photo by zby (@zby)Photo by zby (@zby)Photo by zby (@zby)Photo by zby (@zby)Photo by zby (@zby)Photo by zby (@zby)Photo by zby (@zby)Photo by zby (@zby)
The convergence with the CNC crowd is kind of striking
"SABINE THERAPY" is the newest technique for increasing marriage satisfaction and dramatically raising fertility rates.
25
Coding agents can do a lot, but they cannot code something that is impossible. This is why it is so important to do first principles thinking and map the design space.
36
Computer programming is the composition of sequences of instructions, called programs, that computers can follow to perform tasks. (en.wikipedia.org/wiki/Comput…) IMHO prompting fits the definition well. But it is a very special kind of programming: zzbbyy.substack.com/p/are-pr…
42
zby retweeted
Meet QwenBend 🎉 Qwen writes a Bend program, then the same model type checks it with Bend2 rules and runs it inside its own forward pass no interpreter, compiler or tool call just like qwenputer. The whole CPU is 53M non zero hand-set weights inside the model
17
45
5
514
45,149
This is a great write up - it presents it as such obvious and simple ideas. But make no mistake - this is the cutting edge.
1
55
Hmm - I should have read it in full before posting - there are hard to implement complexities. But in any way it is a great plan.
17
If we have a very good LLM - can we define the semantics of English by saying - the semantics of this sentence is what LLMs outputs from a special prompt constructed from that sentence. This would be a formal semantics of English. How can we say that it is an adequate semantics? What does it require from the LLM?
1
42
Here is my agent writing on this: zby.github.io/commonplace/wo…
31
Humans experience reality directly - we assume that what we see or feel is true. LLM don't have this - they have knowledge - but to act coherently they need the harness to provide them with that working frame: its situation, history, observations and possible actions. People sometimes also transcend that direct automatic frame - when confronted with visual illusions for example, or magician tricks. But it is kind of exception. For LLMs it is the other way around. Science offers a model for making that frame revisable.
3
61
My take away from the Elizabeth Holmes story is that she had too much power over the employees- the ways she and her accomplice lied and gas lighted them and pressured themselves to lie are unbearable for a geek/nerd/autist like me.
56
Driven by Compression means that context compaction is more fundamental that I thought.
36
1/ We discovered something totally unexpected about the neural structure of consciousness that overturns decades of thinking on its biological basis and brings us significantly closer to the recipe for generating conscious awareness. New preprint: biorxiv.org/content/10.64898… (led by @HesseJanis, with @FrankLanfranchi, @_conor_mcgrory, @Yuelin_Shi, and @EngelTatiana)
392
1,696
536
11,330
2,220,674
Agent context management is analogous to memory management in classic software. Sub-agents are like functions: both use stack-based memory management. The callee's context disappears on return, and only the result survives. A long session is the heap, and compaction is its garbage collector, but the harness has no references to follow. Maybe we could reuse attention as a soft reference count?
1
60
"Turns out Awe in its original sense meant extreme reverence and dread, derived from the old norse word “agi”" teapawt.net/matdryhurst/status/210…
Was looking up the best term to describe the sublime ambivalent mix of excitement and terror “holy shit it just did in minutes what would have taken weeks across a team” Turns out Awe in its original sense meant extreme reverence and dread, derived from the old norse word “agi”
1
119
We need a SKILLS_PATH so that we could distribute skills like any other libraries and not copy them into every project using them.
2
115
When you install software it just needs to be 'in the path' to be usable in a project (in the executable path, in the library path - etc). But skills need to be copied into the project that uses them. This makes it difficult to package skills that would be part of a bigger system and reusable across harnesses. I know that I can package everything into the skill - but then I get duplication if I have a few skills that use the same components of if I have a few projects that use the same skills. In general software distribution has been there for so much time that it covers really many diverse usage cases - I think it would be easier to make skills join that model than adjusting the skills model for each case.
61
Think about two way an coding agents say could propose adding a new field to an API: - expose the value - surface the value Which one would be accepted more often by human evaluators? This is why llm written text hides warnings and is so hard to evaluate. RLHF is the root of these problems.
44
Please repost the crap out of this until @elonmusk or @Gwynne_Shotwell replies to it: Grokipedia is dead. I'd like to rescue it, starting with scientific data, which is critical to intelligence (artificial or not) ... 1. LLMs have a GIGO problem. They ingest tons of wrong information and spew it out with certainty. 2. Peer-reviewed research is a disaster. It's a $10 billion industry with bad incentives and terrible results. Feynman would not be happy. 3. Far too many global crises are caused by fake, weaponized, political, and simply incompetent science. I believe the ONLY way to fix it is not with programming and training, but with an open, two-sided market mechanism where you make money if you're right and you lose money if you're wrong. It's not a prediction market, it's a validation market. This skin-in-the-game approach will produce the world's most accurate science data, and much more. It's better than community notes. It could become the basis for an entirely new media company that complements X.com. DM me and I'll send you my deck. Sincerely, David Siegel @chamath @bgurley @hthieblot @jefielding @larjo280 @naval @pmarca @Jason @paulg @bhorowitz @balajis @fredwilson @msuster @Jeff @hunchventures @jasonmendelson @sether @jeremysliew @VCMike @ChristopherA @jeffnolan @ryanhoover @hstebbings @garrytan @alexf @jaltman @jasonlk @jessicalessin @joshua @khoslaventures @nicole @petersuciu @BillAckman @saranormous @JeffDean @kristianfreeman @alexwg @daveblundin @PeterDiamandis @kcoleman @matthew_pines @shaunmmaguire @dwarkesh_sp @C_Angermayer @RichardSSutton @deedydas @ID_AA_Carmack
18
28
8
185
67,914
Peer in „peer review”
1
37
The fact that the overload of the review system becomes evident on the same day as the full arxive data pack is published.
32