Jamie

@RakheeLb @DanielMiessler Not for quantity but for exception/escalation for edge cases. For variance in load quantity, humans are the worst.

@DanielMiessler Real world systems have variance that pure automation can’t absorb. Humans aren’t the scale strategy, more like the elasticity layer when reality deviates from the model. Load balancing is resilience.

Agree it's interesting, but I don't think it's correct. There will be people doing PM-like things of course, but that's turning into a skill vs. being a role. Same with design, or the ability to build/code. And being able to pitch the vision, sell it, etc. And ultimately this is what companies will look like: Graphs of Algorithms. Transparent. Optimizable. Harnesses are what humans use. The core of work being done in companies will be through the execution of SOPs, and they'll look more like the /workflows that Anthropic is about to release. A series of clearly articulated steps that do _________. Humans will definitely have harnesses, but a human using a harness to do something is like an issue that needs to be resolved. Why wasn't that work part of an existing SOP or process? Can it be added? Humans doing things will largely be optimizations of these workflows according to their keen human understanding of the problem and how they want it solved, and coming up with net-new things. But humans doing the main, anticipated work will be a failure case to be solved.

Well said. The saddest part is that he's technically not wrong. A: "Humans are unreliable" B: "A checklist helps a lot" A: "I meant humans without pen and paper and I maintain they ARE unreliable. I was right." B: "This doofus thinks people can't do anything." A: "You haven't even refuted my claim! I am still right!"

Gary, let me be both real and charitable with you: - You are clearly smart and knowledgable about AI, and have a long history with it that people should respect - You seem to care about having AI work out to be good for people in general, and seem to be worrying about the right things - Speaking just for myself, the problem I have is you seem way too focused on being right about LLMs (narrowly) NOT being the future of AI - The problem with this position, which I and many others have repeatedly said, is that there was never a realistic scenario where the labs, or ANYONE for that matter, would build an AI system using ONLY LLMs - Of course there's RL involved. Of course there are symbolic components in the mix. Of course the overall system uses loops and all the existing conventions of programming to create a competent, "intelligent" system. Completely predictable. - Massive tech innovations like this materialize as SYSTEMS, not as a unique, specific, and narrow tech. Even in 2022 ChatGPT required a giant stew of hundreds of technologies to function as a competent application - You're basically arguing against a Strawman version of an argument that the vast majority of people aren't making, which is the idea that a raw neural net, on its own, with no RL or any other tech used as augmentation or enhancement, can function as a full AI system...and that all that's needed is to keep scaling that... - Nobody working with this stuff practically and actually building things is living in this highly pedantic and academic world. In what I would call the REAL world, AI is massively improving every few months, and it's largely because the models keep getting better. Is there a lot of supporting tech around the LLM itself? Yes, but from everything I know at least, it's still mostly the larger/better models (which are still LLM-based) doing the heaviest lifting. In short, the reason you're getting so much hate is because LLM-based AI has continued to improve at an insane pace, while you and others have guaranteed that it could not and would not. It's not a defense to say, "Yeah but they're including for loops and stuff, which is where the real magic is at...", or similar (not a real quote but you get my meaning), because no technology of this significance is a single component. It's always a system. My recommendation is to just be like, "Yeah, surprised me how far it went...good one...but this is really more of an AI SYSTEM than a pure LLM....but I do accept the point of LLM systems in any form being a lot more competent than I thought they could be. And then you can also say, along with the others, that there will be EVEN MORE innovation if we add ______ type of AI, and ___________ type of AI as well, which does ________ better than LLMs. And then people could move to do that. But at this point you're further digging into a pit that basically said LLMs won't get us anything CLOSE to what we have, and yet here we are.

@yoheinakajima Easy: Meta divests manus in exchange for a controlling interest in tiktok

in case anyone is confused like i was, the acquisition did go through, and the order was made to wind that back manus needs to raise and buy back, or it will likely go into some divesture process which means the VCs who were in manus do not have to return the capital they got from the acquisition (plz correct if i’m wrong here)

@GeoffreyHuntley @zex_exe 💯 Correctness proofs are going to be a core mechanic. The ultimate backpressure, not tests. Especially when people get serious about security.

@zex_exe haskell but not good enough to build whats in my head. if i had one last language with a gun to my head. haskell.

home sweet home. time to do a brain dump into a blog post before SF starts speed running the unhinged concept have been pitching and cryptically tweeting. alright if we zoom time back far enough. you’ll end up with the attached. the family tree of operating systems. it’s important to consider this. there was once a time when solari, irix, hpux, aix existed and ruled the world. eventually operating systems converged. this family tree of operating systems converged. now we only have linux, macos and windows. now you might be wondering why i’m hammering on about operating systems. it’s because i’m actually talking about PROGRAMMING LANGUAGES. programming languages are due for a similar convergence event. a programming language (which also might double as a unikernel distributed operating system) that is designed for agents first. programming languages are specifications for machines. markdown is not a programming language. my hot take is a product factory, loom, will not be possible to be solved properly until a convergence event. it’s also why i think a software factory will not be possible via gas town or many of the offerings coming to market this year. face it - programming languages have been specifications of machines but designed for humans. this can be falsified. the adoption issues that developers and programming language authors hold dear can be systematically falsified. we need a programming language designed for agents, for machines. it doesn’t been to be understandable by humans but it does need to be explainable by a machine to a human (via a prompt) so i’ve been enumerating through eso langs and exploring what this could perhaps look like. the training data problem is solvable and my hypo is the convergence event could happen within a couple years. not 10 years. if this seems bat shit insane. well, it is but within my head it’s all logical and want the steps needed to make it happen are pretty clear. not all languages that exit today will exist tomorrow. for example we no longer use solaris and in the distant future i deeply believe dynamically typed languages are solaris and we’ll cease using them.

@steipete @hopsec_ @AnthropicAI I use a tampermonkey script to scrape it from the page in a browser. I avoided cli because /usage will start your usage window, whereas the browser reports it passively.

@hopsec_ @AnthropicAI I open the cli and parse the output, no API tricksery.

Pour one out for the little guys. 😢 Just got a response from @AnthropicAI that my UsageScout app I built cannot use the /usage API endpoint that is used by multiple other usage apps (like @steipete's CodexBar app). #Anthropic #Claude #CodexBar

@Jason I subscribed when it was $3 but now you let all the poor people in and it's gotten so much worse.

Donate $1 a month to charity to roast me (and get 50x the views) in the comments... Memes, dunks, spam... It's all allowed in the comments if you give money to charity!

@mattpocockuk I'd be interested to hear Jeffrey Emanuel's thoughts @doodlestein . Maybe some circumstances or some properties make a difference? (I don't have a take for or against, just trying to learn.)

Long skills are such a red flag to me - Hard to audit (and therefore, trust) - Hard to edit (more text, harder to maintain) - Expensive to run (more text, more tokens) The shorter the skill, the better IMO

@mattpocockuk That was fast x.com/i/status/20546…

@mattpocockuk Who is writing the tmux wrapper to drive claude "interactively" from CLI? Or does it exist already?

This is the clarity we've been crying out for. But it's a poisoned chalice. This is a 10X cut to claude -p disguised as a monthly bonus. Anthropic is discouraging any kind of programmatic usage. And that's fine - no subsidy lasts forever. But it's time to try Codex.

@BretWeinstein @HustleBitch_ @grok how do yamanaka epigenetic effects relate to telomere length

@HustleBitch_ This is nothing like a major breakthrough against senescence. Maybe useful, but mostly hype—as is standard operating procedure in anti-aging research.

🚨 SURGEON WARNS THE BIGGEST BREAKTHROUGH IN HUMAN HISTORY IS HAPPENING RIGHT NOW — “BIGGER THAN AI” A surgeon is going viral after claiming scientists may have already discovered a way to partially REVERSE aging at the DNA level… and he says the implications are bigger than AI, social media, smartphones, or even the internet itself. Dr. Buck Parker’s claim: “The fountain of youth has been discovered.” “This is bigger than the Industrial Revolution. Bigger than the advent of the internet. Bigger than Amazon, Apple, the iPhone, Google, social media… bigger than AI.” And according to Dr. Parker… it’s already happening RIGHT NOW. The core claim revolves around something called “Yamanaka factors,” proteins discovered by Nobel Prize-winning researcher Shinya Yamanaka that can reportedly reset damaged cells back to a younger biological state. According to Dr. Parker: • Scientists have reportedly reversed visible signs of aging in animals • Wrinkled skin in test subjects appeared to become youthful again • Researchers are now experimenting with literally “turning back” cellular age • Human trials are reportedly beginning • Some scientists now believe aging itself may simply be accumulated DNA damage His warning: “If you’ve been alive for the last 40 years… you’ve seen some wild sh*t happen. It’s about to get more wild.” If this became available tomorrow… would you actually take it? 📹: drbuckparker

@allTheYud Also Hubert Dreyfus in What Computers (Still) Can't Do. A classic I very much enjoyed back in the day.

The human brain is a giant kludge. We knew that already. When it comes to seeing that giant messes can do cognition, LLMs add nothing. Some have claimed I am salty about LLMs refuting my ideal of clean cognition. They are MAKING. SHIT. UP. There has long been a debate about to what extent the messiness of HUMAN MINDS implies that we shouldn't bother looking for any underlying principles of intelligence that are simpler than a whole human brain. Eg: If there exists any compact theory of cognition, why isn't there a short program that implements intelligence; and then why didn't natural selection find that short program earlier, and scale up brains to human sizes much earlier in evolutionary history? To this the observed messiness of LLMs adds very little, at least if you just go on the naked fact that LLMs are messy. Human brains are messy too and we already knew about that. There was SEPARATELY a debate, which started around 70 years ago, about whether particular very simple logic-processing programs then called "AI", were stripping away far too much complexity for that entire line of research to ever possibly work out. After all, human brains weren't like that. People who said "I bet that will never work" in that second debate, who were skeptical about the program then known as Good-Old-Fashioned AI, included Douglas Hofstadter, who was a major influence on me; and who was influenced by Drew McDermott's "Artificial Intelligence Meets Natural Stupidity", which I looked up and which also had a big influence on me. At several points in the Sequences I argue for this anti-GOFAI position -- though I had little new to add to what my seniors had already said, on this particular topic. I went over it anyway, because I thought that understanding why GOFAI did not work as cognition, and had been overestimated by its inventors, could illustrate some useful general lessons about cognition and invention. There was likewise a debate about to what extent the failure of the GOFAI program, weighed on whether there was any point in looking for principles of intelligence simpler than a brain. The ongoing failure of GOFAI had ALREADY BEEN OBSERVED by 1996 when I showed up, and I and many others had already updated on it; albeit Robin Hanson in 2008 still considered Cyc a running concern. But the notion of a "principle of intelligence" that does NOT map onto a tiny logic program is MORE ABSTRACTION than IDIOTS ON THE INTERNET CAN HANDLE. I am tired and I do not plan to try to re-explain in detail the difference between a "principle of intelligence" and a tiny little logic-processing computer program. I know that THEY are too bad at handling abstractions to ever understand that. If you don't already understand the difference? Then forget for now that there was ever a debate about something called "principles of intelligence". We were debating about the Foograh of Murbledinks and there is no need for now to understand what that meant. You just need to understand that what we were debating was NOT the little programs processing suggestively named LISP tokens using various logic variants. That debate had nothing to do with those little programs. I was, before LLMs, already part of the position that said, "Those little logic programs are the wrong path and they will never work." Also I happened to take a different complicated nuanced position on an entirely different issue, which may be too abstract for someone to ever understand if they were confusing that other issue with little logic programs. Someone confused in that way will is entirely unable to relate that other issue to LLMs. They cannot operate abstractions of that level to see what those abstractions predict or forbid. You should ignore anything someone tries to say about LLMs refuting something they derogate as "clean principles of reasoning" or whatever, if they are confusing the existence of an explanation for why reasoning works, with little logic programs. One need only remember: When it comes to messiness or NOT being a clean logic program, LLMs add no marginal evidence beyond the prior observation of vastly complicated and messy human brains. That messiness was known and understood and processed into the real debate about that other issue; and had been observed long before I came on the scene. The evidence of human brains was already included in the conversation about that other abstract thing, among the subfactions of that faction that were sensible enough to shrug aside the little logic programs; people like McDermott and Hofstadter, and much later me. Perhaps somewhere on Earth there is a fool who is salty about how messy LLMs succeeded, because *they* thought it was going to be a neat program processing suggestively-named LISP tokens using a clever logic variant. That person is not me. I talk about the folly of the little logical programs in the Sequences. In the Foom Debate I argued to Robin Hanson that handcoding large amounts of knowledge into a logical representation framework like Douglas Lenat's Cyc was not a promising project. The people saying I endorsed the little logic programs MADE THAT SHIT UP. For that sounded cool to them and like it ought to be true, and they were ignorant of history and bad at abstractions. Call it the Logic Libel.

@greg16676935420 @TickPick Superman 3 extra half cent fee: $0.00*

@TickPick StubHub be like: Concert ticket: $50 Venue fee: $26.10 Access fee: $14.28 Paperless transaction fee: $7.23 Convenience fee: $17.24 Fee fee: 6.23 Fee Fi Fo Fum fee: $9.23 Because we can fee: $8.12 Another dollar won’t hurt fee: $1

Thought this was StubHub checkout

@GaryMarcus @ChrSzegedy Is this even bad? I like that you are out of distribution.

@ChrSzegedy honestly this cheap shot is beneath you, Christian and the truth is that many many people have moved to my side re the critical importance of distribution shift

Some things never change: Gary just doesn't fit in distribution.

@Jason @KritterKS I know you know that in Austin that is called a lake.

@KritterKS Is that a mudslide or a river?

Hot diggity damn - my room in San Antonio has a riverwalk view!

@DanielMiessler @kenswain615 Equally stupid would be collecting by voiding treasury debt that they hold.

My buddy @kenswain615 and I just came up with a sick idea to solve the debt that's stupid and will never work. We make a list of all the IP China has stolen from us, including the recent model distillations. And we send a Stripe invoice for "N Trillion" dollars. Debt cleared.

@allTheYud A: What would you say if you were bad at counterfactuals? B: But I'm not bad at counterfactuals. A: Right! B: What?

People are just so incredibly gullible when it comes to negative claims about rationalists, in reality the largest group of organized one-boxers on the planet. One-boxing as signature move. 2.7K likes for MAX IRONY.

@liron @BrandonGoldman We need a new category for disability, meaning people basically unemployable through no fault of their own. Disabled is any IQ below 90. Oops now 110. Oops now 150...

I claim (60% confidence) that 2 years will be enough time for data to show an early trend of AI pushing humans permanently out of jobs! So I’m happy to bet with Will at 1:1 odds. Yes, I already lost a similar bet in the 2023-25 timeframe against @BrandonGoldman, but I persist.

@GaryMarcus What makes the difference between "AI" (neuro-symbolic or otherwise) and a dumb wrapper or harness? Some hard-coded string matching in a LLM wrapper is enough to be neuro-symbolic AI or is it something else?

Claude Code is not AGI, but it is the single biggest advance in AI since the LLM. But the thing is, Claude Code is NOT a pure LLM. And it’s not pure deep learning. Not even close. And that changes everything. The source code leak proves it. Tucked away at its center is a 3,167 line kernel called print.ts. print.ts is a pattern matching. And pattern matching is supposed to be the *strength* of LLMs. But Anthropic figured out that if you really need to get your patterns right, you can’t trust a pure LLM. They are too probabilistic. And too erratic. Instead, the way Anthropic built that kernel is straight out of classical symbolic AI. For example, it is in large part a big IF-THEN conditional, with 486 branch points and 12 levels of nesting — all inside a deterministic, symbolic loop that the real godfathers of AI, people like John McCarthy and Marvin Minsky and Herb Simon, would have instantly recognized.* Putting things differently, Anthropic, when push came to shove, went exactly where I long said the field needed to go (and where @geoffreyhinton said we didn’t need to go): to Neurosymbolic AI. That’s right, the biggest advance since the LLM was neurosymbolic. AlphaFold, AlphaEvolve, AlphaProof, and AlphaGeometry are all neurosymbolic, too; so is Code Interpreter; when you are calling code, you are asking symbolic AI do an important part of the work. Claude Code isn’t better because of scaling. It’s better because Anthropic accepted the importance of using classical AI techniques alongside neural networks — precisely marriage I have long advocated. It’s *massive* vindication for me (go see my 2019 debate with Bengio for context, or to my 2001 book, The Algebraic Mind), but it still ain’t perfect, or even close. What we really need to do to get trustworthy AI rather than the current unpredictable “jagged” mess, is to go in the knowledge-, reasoning-, and world-model driven direction I laid out in 2020, in an article called the Next Decade in AI, in which neurosymbolic AI is just the *starting point* in a longer journey.* Read that article if you want to know what else we need to do next. The first part has already come to pass. In time, other three will, too. Meanwhile, the implications for the allocation of capital are pretty massive: smartly adding in bits of symbolic AI can do a lot more than scaling alone, and even Anthropic as now discovered (though they won’t say) scaling is no longer the essence of innovation. The paradigm has changed. — *Claude Code is plainly neurosymbolic but the code part is a mess; as Ernie Davis and I argued in Rebooting AI in 2019, we also need major advances in software engineering. But that’s a story for another day.

@eshear @tautologer If you were to replace "final goals" with "values", would you hold the same claim?

“The first, the orthogonality thesis, holds (with some caveats) that intelligence and final goals (purposes) are orthogonal axes along which possible artificial intellects can freely vary—more or less any level of intelligence could be combined with more or less any final goal.” The problem with this claim is that there final goals are a thing that exists. In reality there are only shorter and longer term goals. A final goal is a longest term goal, but there is no longest time horizon. Only the longest we’ve understood so far. Could I persuade Ted Bundy? Almost certainly not, but that doesn’t mean anything. I can’t convince most people of most things. But that’s just because convincing ppl of things via rational argument is really hard in general.

people who don't believe the Orthogonality Thesis: do you believe that there is a rational argument that you could present to Ted Bundy that would convince him that killing people is wrong, such that he no longer was motivated to do it?

@ewarren Good luck with apportionment: "direct Taxes shall be apportioned among the several States which may be included within this Union, according to their respective Numbers..."

The Ultra-Millionaire Tax Act would generate over $6 trillion over the next decade—without raising taxes on 99.85% of American households. This wealth tax for millionaires and billionaires could pay for universal child care, free community college, Medicare expansion, and more.