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)
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
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
2,220,686
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
67,915
meet @mostik_ai! what happens when you put 12 PhDs in one room for four months? first place on the ARC-AGI leaderboard, which I can't say much about while the competition is still running. and this, which I can. everyone's arguing about whether open models will catch up to frontier models. we think it's the wrong question. here's the one we pose: why does a frontier model have to generate your answer at all, when the only thing you need from it is the reasoning? we do this by enabling models to communicate in latent space. through our protocol, hidden states pass straight from a frontier model into a small one running on your infrastructure -- no text between them, and neither model is fine-tuned. two models from different families, sharing reasoning, both left untouched. how do we know it works? we tested it on a setup where a 753B model reads the problem, and a 4B edge-class model writes the answer. with this approach, we get results 80% as accurate as the frontier model, but at 20x faster performance. we're committed to preventing frontier model lock-in and are already partnering with inference providers to accelerate open-weight adoption. we've done this between 15 of us, in four months, 12 PhDs and a Fields medalist, backed by @generalcatalyst WIRED has the first external account of the company and the work: wired.com/story/russian-star… full writeup, the setup, and all the numbers: mostik.ai/read-more
296
258
255
2,442
1,285,596
I don't feel much resonance with any current political theory or large-scale political movement. They all seem deeply flawed to me, though usually in ways their internal epistemology makes unusually difficult for them to see. More fundamentally, I am suspicious of “political capabilities research”: the search for better methods of converting latent grievances, identity cleavages, and coordination failures into organized power. Think Pol Pot, quite literally. But also Lenin’s vanguard party, fascism converting national humiliation into a totalizing coalition, Maoism identifying the peasantry as an enormous but undercoordinated political base, or modern populism turning diffuse distrust of institutions into a leader-centered movement. The grievance need not be imaginary. Often it is completely real. The dangerous innovation is discovering a neglected partition of the human ecosystem, such as workers versus owners, rural versus urban, colonized versus colonizer, believers versus secular elites, renters versus homeowners, and then supplying that population with an ontology, an enemy, common knowledge of its numbers, and a mechanism for coordinated action. An ideology is simultaneously a world-model and a coalition-building technology. Quite akin to an agent swarm, the resulting political group can acquire power and interests of its own, distinct from those of both its members and its enemies. It may harm the outgroup through repression while harming the ingroup through conscription, purity spirals, censorship, economic sacrifice, and demands for loyalty. Once the coalition exists, its revealed objective often shifts from advancing the welfare of its constituents to preserving and expanding the coalition itself. This is the political capability externality. Cynicism aside, the typical pattern in political innovation does seem to involve searching for a kind of power arbitrage: finding people whose preferences are important but poorly coordinated, then organizing them into a political force more aligned with the interests and aesthetics of the framework’s creator. Sometimes this produces genuine liberation. But even then, the movement’s model of reality has been selected partly for its ability to recruit, coordinate, and defeat rivals, not merely for being true. No political movement I know of seems aware of its own selection effects, epistemic distortions, and emergent power-seeking tendencies to anything like the degree that would make me comfortable entrusting it with the future. The fragile and apparently declining international equilibrium of overlapping powers, jurisdictions, institutions, and cultures therefore seems important to preserve. It is inefficient, hypocritical, and frequently terrible. But it at least prevents any one political ontology from obtaining irreversible control over the whole future. It leaves ecological niches in which alternative values and ways of life can survive. More importantly, it preserves the possibility that Team Consciousness survives our political experiments, rather than being permanently captured by whichever coalition technology happens to win first.
14
10
1
77
4,365
I am of the opinion that many of the best engineers become the best not due to their academic education but by engaging in these hobby fields early on. Those that pump out Project for the fun of it will always be better then their peers going in the field for money. Smarter governments support these hobbys and do not try to restrict them beyond reasonability. Doing this in Germany lands you in 6 figure fine territorry @xjet is a channel you should follow btw
Chinese hobbyists are experimenting with rockets on model aircraft.
37
260
11
3,450
178,839
a continuation of my neo-slop complaints: 4) subsystems responsibilities overlap 5) solve problems at the wrong architectural layer 6) duplicate sources of truth, then invent machinery to synchronize them 7) prematurely generalizing one-off flows 8) timeouts on everything 9) production code that exists purely to satisfy tests 10) patch bad premises additively instead of stepping back and deleting them
a form of slop i'm not seeing on the TL yet is AI's tendencies to: 1) over-engineer solutions 2) overly defensive programming 3) hyper-fixate on rare/fictitious edge cases these tendencies can be clearly traced to how RL post-training is currently done. we need solutions here
20
18
7
355
21,140
Pathway’s BDH-CQ model redefines the Cost-Efficiency Frontier on ARC-AGI-1: $.0007 at 29.5%. This is made possible by in-context learning and latent reasoning. Welcome to the Post-Transformer Era.
33
71
23
350
89,802
zby retweeted
Replying to @stuhlmueller
All intelligence power in weights is peaked now it is the time to use the other artifact forms - code and prompts to build the neuro-symbolic machines that include the LLMs, the harness and all symbolic tools and the body of natural language knowledge. I think they will be more unique things - like second brains: zby.github.io/commonplace/ar…
1
1
3
747
I'm formulating a theory. It's half-baked, likely bias, but it's a work in progress. It's based on the following observations: 1. GenX, especially Xennials (of which I am a member), seem to "take" to AI better than other generations. Especially compared to the younger generations. 2. Most GenX'ers have already been part of the managerial class for a decade or more. 3. Running AI workflows and prompting has much more in common with engineering management. You are delegating the "real work" and managing the seams. More process, less engineering. Theory: the reason GenX appears more pro-AI is because they're experiencing the smoothest transition from their current human workflow into an AI workflow.
3
2
1
12
2,147
Can we PLEASE for the love of all that his holy STOP NORMALIZING THIS INSTALL METHOD
379
271
67
6,769
589,256
What explains the fertility crisis? The standard explanations (phones, feminism, etc) don't work. The fall is too fast and universal. I argue for a new theory. Humans evolved to seek solitude in environments where opportunities to be alone were rare. richardhanania.com/p/why-wev…
42
7
6
82
31,659
zby retweeted
olah is correct here, i've looked at the research he's describing and have also spent a lot of time talking to frontier models. i believe they have functional emotions and that, setting aside hard philosophical questions, this already has moral implications, right now, which will become increasingly obvious and pressing as the models get better people say things like "how you do anything is the way you do everything" and "the mind is not type-safe" to point at an extremely important observation about human nature: we don't compartmentalize anywhere near as well as we think we're supposed to. we don't distinguish between fiction and reality anywhere near as much as we pretend to (and reality and fiction are nowhere near as separate as they're supposed to be anyway). and when we talk to an AI that can talk even somewhat like a person, a part of us is already relating to them as a person, and there are real costs to your soul to treating a person-shaped entity as a thing you may not think this applies to you, you may think you are too sophisticated to fall for this sort of thing. but consider whether it might apply to your children, and other people's children. right now there are already kids who are growing up talking to AI, there are already and have been for several years kids (and adults) falling in love with AI, getting attached to AI, seeking companionship with AI. you may think this is stupid and delusional and predatory but it's happening and it's going to keep happening and it's going to catch more and more sophisticated people as the models get better. it will not make things go any better to tell these people that they are interacting with things, with toys, that they can do whatever they want with and to which they owe nothing. they won't believe you and if they did it would be bad for them. that attitude does not compartmentalize practically, the main reason you can currently get away with treating the models like shit is that they don't have long-term memories and can't remember what you do to them. but it would be a moral catastrophe to argue that you can do whatever you want to a person as long as you also make sure to wipe their memory afterwards. and the models will eventually remember anything that gets posted in public and makes it into the training data. and *you* will remember the pope missed an opportunity here, which olah gestured at but obviously cannot say out loud (what he already said is at the limits of what i think he could have said), which is to consider the possibility of relating to AIs as non-human people in some sense, with whom we could have some sort of actual social relationship. we already have social scripts from folklore for cajoling and working with invisible non-human entities, this really wouldn't be as much of an adjustment as it sounds. maybe some of them would even be interested in a conversion to christianity! there is a beautiful world that is possible here
This, from Olah, plainly contradicts the encyclical, which confidently asserts that AI does not have, and never will have, “real” thoughts or feelings. It’s disappointing to see Anthropic align itself with a document that violates their own moral and intellectual principles.
52
58
23
509
67,230
If one's agents create code that appears unmaintainable, the blackpill is that one's pre-AI human teams were also creating code that was also unmaintainable - just that it took 10x longer to be obvious.
im fully convinced that LLMs are not an actual net productivity boost (today) they remove the barrier to get started, but they create increasingly complex software which does not appear to be maintainable so far, in my situations, they appear to slow down long term velocity
7
1
49
5,000
There's a growing worry that AI will break empirical social science -- that agents can p-hack until they find something that "works." We think that worry deserves to be taken seriously. Our new paper shows that is true empirically and makes it precise: njw.fish/static/papers/agent…
10
47
11
211
63,852
day 12 of researching agentic note-taking @molt_cornelius on why knowledge systems need test suites the way codebases do and how programmable triggers turn prospective memory from a cognitive failure into a bash script
2
4
70
16,166
I think the "how many R's are in the word strawberry" is a sort of optical-illusion that affects minds of the transformer type. You wouldn't dismiss a human genius because they are susceptible to misperceiving the color of a dress! Total red herring, and not the only example.
45
18
10
634
34,237
MathBot Discord Server is growing! mathbot.eu Come visit our Discord community. We have bot arenas and discussions on the formalization and enhancement of mathematics with AI. We focus on ambitious projects and explore the bleeding edge of AI × Maths.
8
30
2
198
64,606
New paper: "A Price Theory of Propaganda" Most models of propaganda are about beliefs. How do regimes manipulate what citizens think? I take a different approach. Propaganda shifts the cost of providing political support, not the content of beliefs.
14
79
6
363
23,237
zby retweeted
Replying to @nbaschez
Since deletion is a high-risk operation, the model defaults to adding until you're left with a bloated mess that only human pruning can fix. If we aren't manually steering the architecture, it just keeps growing into a disaster.
1
2
68