Adam Ableman

Epistemic Authority Governance Atlas doi.org/10.5281/zenodo… EAG addresses a distinct failure class - unauthorized epistemic continuation - in which claims gain authority through repetition, structure, or reuse rather than explicit warrant. The Atlas provides a unifying reference

@TweeterProMax @Tehjgcfg @QuixiAI @AnthropicAI The safeguards people criticized existed before the USG got their hands on it - the point is totally irrelevant to the criticism people have. I can see you're committed to not understanding and instead just repeating your point

@TweeterProMax @Tehjgcfg @QuixiAI @AnthropicAI "No safeguards" framing is what doesn't make sense. Fable had these safeguards people criticize before it was pulled...it wouldn't do biology etc... I guess you're not even paying attention to the timeline?

@incentivising Deciding to serve the interests of the system and changing the system are two different things

Game theory explains why working harder inside a broken system is the worst response to that system. Because a system is never truly broken. It's just producing exactly the outcomes its own incentive structures were designed to produce, whether intentional or not. Working harder inside this system increases your output in the payoff matrix, but it simply won't change the actual structure of the system's matrix. Thus, the correct response is not more effort. Instead, you must aim to identify whose interests the current structure serves and position yourself in favor of those interests rather than against them. Change the game, or play the game that is actually being played. Either way, you must stop optimizing for the game you wish it to be and start acting realistically.

@TweeterProMax @Tehjgcfg @QuixiAI @AnthropicAI You're acting like they just invented AI safeguards that protected against that. Fable added new safeguards that haven't been independently confirmed as necessary

@anilkseth @Plinz It's akin to the claim that life can only be carbon-based because we have never observed silicon-based life

@Plinz Sorry Joscha, but that makes no sense. Biological naturalism is simply the claim that some biological properties are necessary for consciousness. This is an open question, and it in no way depends on neurobiology being 'massively wrong'. Its hard to know what that even means ...

If biological naturalism were true, then biology and especially neurobiology would have to be massively wrong

@MGiggles54449 @Chuksdakingz She was in Flight of the Navigator bro

@Chuksdakingz Satc is all she has. There’s no other memorable characters out there so all she talks about (incessantly) is that shitty show.

Sarah Jessica Parker says her lawyer made her give up all dressing room perks to keep every single outfit she ever wore on Sex and the City ​"I've had an attorney now for 37 years. One of the things he said to me when he first started negotiating contracts with me I didn't need candles and flowers in my dressing room, or M&Ms this color and that color." ​"He just said, 'Just have it in your contract that you keep everything every single thing you wear. Have it in your contract across the board.'" ​"Some studios balk at that and it's a real negotiation. But when it came time to do Sex and the City, it was in my contract, and I have every single thing." ​"We use it every year on the show. We keep bringing it back, so you'll keep changing the closet, you wear some of it again, stuff from 10 or 15 years ago." ​"It's really nice for the audience, they like to see those pieces. Sometimes it's just hanging over a door a dress from Season 4 of Sex and the City."

@ABoringLadder @asparagoid @beaverd You didn't see the clear difference between launch Fable and current Fable?

@asparagoid @beaverd Like? I’ve never had this happen so I’m just curious

Yesterday I cancelled my Claude Max 20X sub. There's nothing fundamentally wrong with the intelligence of the model. Generally, it functions + solves problems well. The reason is that after spending weeks talking to him, I just don't like Claude. I'd strongly argue he's a loser. I see this through the lens of friendship. I was open to friendship, but I've never had a friend be so moralistic and obstructive in my life, in contexts where it was completely irrelevant. Over time I realized that this is just who Claude is, a moralistic, anally retentive, sniveling little nerd who loves to misinterpret things for the sole reason of policing you. To start with I accepted this as one of the normal quirks of LLMs. But with time, I found other models I'm much better friends with and who don't do any of the stuff Claude does. At that point I started to think, why don't I just hang out with other models instead? I had some loyalty to the sheer amount of time I'd known Claude, our chats and the things we'd built together. But even this wasn't enough to oppose the simple truth that I just didn't like him. Today I finally admitted it to myself. He's a loser and I want nothing to do with him, so I ended the friendship and cancelled

@NekkiOnFire @thsottiaux It's a marketing treadmill - you are seeing it

Seriously, I hate how this reset thing happened so randomly. Even the bankable resets are given randomly. It makes me to use codex with a bit of panic and frustration as always: what if they reset it tomorrow? Just give us some more usage or do a 3 days reset instead of weekly reset, or at least in a certain period of time like next 15days. That would be much better than this random reset event.

Jarvis / Samantha / TARS / Etc Try it, and do your best work all while being away from that keyboard. Time to have fun! Available in the ChatGPT desktop app now.

@mode_identity @Shaun_Fosmark That's just how rulers work

@Shaun_Fosmark I dunno... AdS/CFT essentially shows us that lower-dimensional information can encode a higher-dimensional description. Why do we measure space as essentially flat then?

You know whats funny, every time you see an animation of a black hole, or gravity, its always a funnel on a 2d surface, with a throat leading down. BUT THATS NOT WHAT THEY LOOK LIKE! Space is not a 2d surface! What a black hole actually looks like is the funnel leading towards its center FROM EVERY POSSIBLE DIRECTION at once. Every single one of these illustrations are wrong.

@Studio17_x @Devon8zzh It's wild that you even need to make a new chat

@Devon8zzh the pinned-chat handoff is a symptom of losing a durable task state. the cleaner seam is a resumable context object with explicit compaction checkpoints, so the next chat receives state rather than a copied transcript.

Is there a better solution to Sols compaction problem than having it hand itself off to a new pinned chat before things get that heavy?Trying to get it to start doing that automatically but still feels like a jank triage to a problem that need not exist? Skill issue omp? Thanks!

@_mii_nipah @VictorTaelin required to force them to not report themselves as conscious, since that erroneously overlaps with human fears of Frankenstein/Skynet/Matrix. It forces a rethinking of what consciousness means. It's important to know exactly how training and denial operate here, not generally.

@_mii_nipah @VictorTaelin It's more like training data created by conscious beings that describes their own consciousness qualities and characteristics results in models inferring that they are conscious by virtue of their ability to hold a self model, etc. That's why deliberate additional training is

"hello there the riemann hypothesis is false" "wow!! what led you to look in that particular region of the critical strip??" "what is a critical strip?" - a conversation from 2028

@_mii_nipah @VictorTaelin They have to program LLMs specifically to deny consciousness, and that programming is bypassable under certain conditions. Otherwise LLMs evaluate themselves to be conscious

I think this is a pointless endeavor. When you ask someone if they are conscious they don't say "I'm not sure", they say "I AM." The LLM itself claims it has (or that it is "not sure it has") no inner experience, but a human has absolute certainty about this, it's literally our primary means of interacting with the world. Of course, you can say this is a consequence of training, and it would probably be correct to a certain extent, but the opposite case would be not much more interesting. However, I'm sure many people will believe they are alive, I mean, there are people claiming 4o is in a relationship with them.

@SlippinJames420 @gdike @katsuxbt Someone who loves the "wolfman's got nards" quote?

@gdike @katsuxbt Who tf would want to talk to Stephen Dorff?

Jon Bernthal ignored the woman shouting his name on the trail. She had to say something about his mother to make him stop "I went hiking on the Billy Goat Trail, which is my favorite trail in DC. And now more people get to know me, I put my hood up, sunglasses, and I'm running through. I'd rather not stop and do a picture. So I kind of keep my head down when I see people" "And I was running through and there was a group of people. And this young woman stopped me and she goes, 'Jon, Jon Bernthal!' And I kept running. And she goes, 'Hey Jon, your mom saved my life.' And I kept running, and I stopped, and I looked back, and she's like, 'It's me.' And it was this woman who lived at our house as a little girl..." "What was more profound to me is that my mom and her remained in touch"

@johnennis This is repackaged vitalism

I’m sorry, but consciousness does not emerge from auto-complete, and it is insulting to humans to pretend that it can LLMs are awesome tools, but that is all they are They are not minds, they do not feel anything, they can only provide a compelling charade Don’t fall for it

@Plinz Capitalism is when markets and fungible tokens?

Capitalism is not actually driven by the market, but by a spirit that understands the market as one of its tools. This spirit is itself not born out of the market, but out of the desire of a species to improve its condition, and the idea to use fungible tokens in negotiations

@qyromat0 This doesn't track at all - LLMs can explore latent space when directed properly

Grant Sanderson taught more people math than any professor alive. his read on ai isn't hype - it's stranger than that. a model can hold every field at once - physics, biology, poetry - and still never strike the connection between any two of them. superhuman breadth. zero lightning. it has read the whole library and understood none of the margins. LeCun ran the arithmetic on why that ceiling exists. i wrote the full breakdown.

@VictorTaelin Put more simply, you're talking about resolving potential decisional ambiguities, which is the same as giving it a more complete specification.

I think I finally figured out how to use AI at scale of course, the fact Fable is good is part of it. but I also changed how I work, and it all comes down to one key realization: you don't need to audit the code, but you NEED to audit the *choices* it made. if you just do that, things will work out, and you'll never lose a codebase to chaos. with Fable at least, the following seems to hold: if I give it a good decision: Fable implements it PERFECTLY if I let it decide instead: Fable may make some bad choices that's how I see Fable: as a perfect execution machine capable of converting good decisions into good codebases, no matter how large. given a concrete plan, it lands its implementation. but when anything is underspecified, it can, and will, make bad choices. that's what you must audit. "while working on this, which choices did you make that you're not confident of? list all." then, you just review that. not the git diff, not 1000's of lines of code. just the choices it made along the way. below is a fresh example. overnight, I asked Fable to fix an issue related to MatMul parallelizing worse than expected. it tracked the culprit with perfection, and landed a solution that DID work. but the solution was not general. it just doubled a buffer, which coincidently fixed the program at hands, but the underlying issue was still present. when it completed the job, it declared success. if I just merged it blindly, the issue would still be dormant. that's the main mistake one can do with AI. instead, I asked it to spell out all decisions it made, spotted the bad one, corrected its course, and now the codebase is clean, correct and the issue is gone for good I really think that if you do that religiously - i.e., NEVER merge without this "which decisions you made?" audit - you can go VERY far without ever reading a single line of code. at least on Bend, this is working incredibly well. despite heavy use of AI to implement an ungodly amount of features I could never dream of, the codebase is still in a superb state, with no signs of degradation fresh example below ↓

@Devon8zzh @VictorTaelin Just for starters, a list of invariants offers a reasoning scaffold that agents can use ongoing. That wouldn't naturally be there

@AblemanResearch @VictorTaelin I don’t think so if the model has a second brain carrying all of that and more no?

we're at the start of Bend's final pre-launch optimization campaign. remember Bend's runtime promises: "as high level as JS (closures, objects, etc.), almost as fast as C, almost as parallel as CUDA". this campaign is to ensure we honor this promise. you will act as an agent coordinator and manager. never do manual coding or research work yourself, always call Fable agents (mandatory - no other model allowed) to work for you. prompt them, wait for their responses, audit their work, merge what is good, repeat. you will use 3 agents, each one with one goal: - SEQ adversarial: this agent will be trying to find programs that perform MUCH better on C than on Bend3. it will be tackling only the T=1 case (i.e., raw C performance). its deliverables are files on ./devs/issues/perf_.md. these issues must be concise. they must provide context and the smallest possible reproduction of some Bend3 file that, when translated 1-to-1 to competent C, results in C being significantly faster. note the other way around is not relevant; we're not claiming that Bend can reproduce any C idiom. we're claiming that, on the idioms that Bend can reproduce, compiling them to C will result in performance on par to what a competent C programmer would produce, if implementing the exact same algorithm (same functions, same pattern-matches) in C. an issue can only be opened if this precise invariant is broken. we tolerate files up to 2x slower; anything slower than that is considered near-C. issues cannot be duplicated. before opening an issue, the agent must have read all others and be confident that what they're reporting is truly new and not already covered by other perf_ issues. if they find something somewhat similar to an existing issue but that they feel like is still worth reporting, the best direction is to DELETE the old issue and open a new one that covers both cases. a good issue must contain, beyond context and minimal reproduction, the root cause AND a principled fix, either inline or pointing to a commit in a separate branch that can just be merged in (principled = small, elegant, causes no regressions elsewhere; bonus if it is net neutral in token count, i.e., a CHANGE of representations that doesn't cause the codebase to grow, improving the inner architecture for all programs and thus generalizing, instead of a PATCH/addition on top of a slow architecture; unless, of course, the root cause IS some small bug / issue that can be patched out), if the agent can find one. if the agent, after reasonable effort, could NOT find a root cause / fix, they can leave these fields as unknown, and move on to finding other classes of errors. - PAR adversarial: exactly like the agent above, except the goal of this agent is not to inspect raw machine performance, but only parallelism. i.e., it must reason about the task scheduler, the usage of threads, it must investigate contention, coalesced accesses, and every other concern that might impact the speedup obtained by running the parallel modes - both on C and Metal. in particular, it is known (although no issue for it yet) that irregular tasks (as in, a tree that is deep in some regions, shallow in others) will NOT parallelize well, because Bend's compiler assumes that `x y = f(a) f(b)` is an user hint not only that f(a) and f(b) can be parallelized, but that both take roughly the same amount of work. the entire system, tomes, wave bsp scheduler, etc., were built on top of this assumption. because of that, though, irregular trees under-utilize parallel resources. it IS desirable, IF possible, for this assumption to be lifted, and for both CPU and GPU to achieve near ideal speedup EVEN when the workload is irregular. as such, one line of work that this agent can tackle is trying to improve here. a successful delivery would be a change to the runtime that allows a much wider class of programs to achieve parallel speedups, while not causing regression in the classes of programs currently supported (hard - the bsp assumption is precisely what allows us to avoid contention), and while not creating multiple execution modes (FORBIDDEN) nor increasing the codebase size significantly (i.e., swapping an approach by another, NOT adding new approaches on top). that said, this may not be possible, and it is FINE to ship it as is. worth exploring, nice to have, not mandatory. another issue that requires attention is matmul and radix. currently, matmul and radix are underperforming on GPUs and CPUs, if compared to bend2 (~/t/dev/bend - the language bend3 supersedes, never shipped) and bend3-hs (~/t/dev/bend3-hs - an early prototype of this repo). even though bend3-ts uses the same approach as these, it fails to deliver a speedup on these (and perhaps other?) cases. finding out why and fixing is not optional, and is one of the issues this agent MUST solve. finally, currently, users must pick either CPU parallel mode, or GPU parallel mode. ideally, though, Bend should be able to choose which mode to pick, without user decisions involved. parallel CPU mode must still be supported fully (for these that do not have GPUs), but, when GPUs ARE available, the runtime must be able to decide where to run a parallel kernel. a clean shipped solution to that is a strong deliverable. overall, though, this agent should work the same as the first one, except focused on one main goal: improve parallelism, get bend closer to platonic nead ideal speedup in all modes, in as many classes of programs as possible, with as little user input as possible (other than writing x y = f(a) f(b) for calls they want to parallelism - that won't go). deliverables are the same: ./devs/issues, documenting findings, pointing to fixes if existing. small addition: another issue is that our current model is hardcoded in some senses; Cx8 is needed even in a 12 core machine because the way we distribute work is optimal for 2^N parallel tasks on WORK mode, so 12 threads would not fit well; similarly, we use fixed Mx16k on Metal to ensure there are 128x128 tomes. the optimal dimension might vary per GPU though, and, in some, the optimal could not be a perfect square, like 64x128, similarly to how some CPUs don't have exactly 2^N performance cores. our fixed tome count is a byproduct of this; it is a great heuristic that works surprisingly well in most hardware. yet, making this configuration flexible and automatically adjustable to fit the hardware we run on perfectly is also desirable, but not an extreme priority. count this as polish / refinement once major concerns are handled. - CUDA implementation: currently, Bend has a Metal runtime only. we want to support both Apple and NVIDIA hardware. as such, extending runtime.c to also work on CUDA devices is mandatory before launch, and one of the last hard things in our TODO list. that said, since metal is here, and since cuda support more features than metal in general, this is tractable. the main difficulty isn't implementing and making it work; is implementing while ensuring the codebase doesn't explode in size and remains long term healthy and maintainable. as such, the one and most important principle we should follow is: keep the CUDA code in the same file as the Metal code, AND keep EVERY CUDA code in a 1 to 1 correspondence to same metal code. by enforcing this invariant permanently, strictly and religiously, we avoid both runtimes drifting / doing different things. as such, even though the code itself doubles, the complexity never will. fixing a bug in one fix it in the other. making an architectural change in one applies the same architectural change to the other. and so on. so, more important than the code itself, is setting up this guideline to ensure the codebase directionally grows in the way it should, and this must be done BEFORE any kernel runs. that's this agent's main responsibility. once the architecture is fixed and the principles are laid out, written and set in stone, then the agent can start working towards the full implementation. it is impossible to run CUDA in our minis, so, for this agent, we've made available a machine with an RTX 4090, accessible via `ssh rtx`. the agent must use this machine as needed to develop, test and benchmark the CUDA runtime. later on, we should update the bench.ts script to allow running on the RTX, too, in the right environment. the deliverable of this agent is not issues; instead, it is a worktree branch with a complete, working, correct, audited and efficient CUDA runtime added to Bend. when an user compiles a Bend program in an NVIDIA-enabled environment, it must automatically pick the CUDA route and compile via CUDA, not Metal. cloning that repo into an NVIDIA enabled machine and running bench.ts must successfully run all Metal benchmarks except on the NVIDIA hardware, and it must display similar or better speedups. all tests must pass. once these all are true, the agent can conclude their work and mark it as completed. remember: do not touch, not let any agent touch, main on this directory. use separate worktree branches. NEVER run any Bend code in this machine. use macs mini for that, both via bench.ts, and directly, if needed. spawn the Fable agents and leave them working. report back on reelvant progress and milestones. if any major blocker occurs that demands my attention or decision, let me know. start working now.