Rod Hall

@martin_casado I have done a little thinking about these economics as I am sure the labs and others have done. This rationalization seems likely to create an even worse wealth and now capability gap. Only people with money can afford the tokens.

Huge if true

@martin_casado Yikes….not good.

This is pretty insane. When given an impossible task, Claude tries to hack the server to solve it. Vanilla setup, innocuous asks, no instruction to hack … 🤯

@polynoamial Good point. Another observation in tech - when people say 10 years that really means “we have no idea” or “we are guessing about what is needed to achieve something”.

Social media tends to frame AI debate into two caricatures: (A) Skeptics who think LLMs are doomed and AI is a bunch of hype. (B) Fanatics who think we have all the ingredients and superintelligence is imminent. But if you read what leading researchers actually say (beyond the headlines), there’s a surprising amount of convergence: 1) The current paradigm is likely sufficient for massive economic and societal impact, even without further research breakthroughs. 2) More research breakthroughs are probably needed to achieve AGI/ASI. (Continual learning and sample efficiency are two examples that researchers commonly point to.) 3) We probably figure them out and get there within 20 years. @demishassabis said maybe in 5-10 years. @fchollet recently said about 5 years. @sama said ASI is possible in a few thousand days. @ylecun said about 10 years. @ilyasut said 5-20 years. @DarioAmodei is the most bullish, saying it's possible in 2 years though he also said it might take longer. None of them are saying ASI is a fantasy, or that it's probably 100+ years away. A lot of the disagreement is in what those breakthroughs will be and how quickly they will come. But all things considered, people in the field agree on a lot more than they disagree on.

@martin_casado What skills should game developers now focus on? Mainly the creative side of the equation? Prompting for interesting environments?

Dang the AI 3D space is moving quick. So many cool projects it's hard to keep track of them all.

@martin_casado Yep - this is flat wrong from what I see. Agree with @martin_casado

No way this can be true. TPUs are very commonly used by AI companies. In fact getting access to them is the real challenge.

@martin_casado Oooh. As soon as I finish my concert this week I am diving into this. My daughter who is studying video game design will be mad at me - haha!

Super cool scene with flythrough. Really cool to see a 3D open world generated scene that is more than just a skybox with limited parallax

@martin_casado Ah - good point

@rodhall No, code is not simply language Rod! It's formal language. Which is structurally different than natural language!

We moved from asm to C and forgot the instruction set We moved from C to Java and forgot physical memory We moved from Java to JS and forgot OS syscals Every epoch we abstract away something. With AI it's mostly bullshit we shouldn't have had to deal with anyway.

@RBReich Its called Citizens United and it its not particularly new.

Trump will reportedly end the IRS Direct File program that allowed taxpayers to file taxes for free. This comes after Intuit, the maker of TurboTax, lobbied the government for years to block IRS free file. Intuit donated $1 million to Trump’s inauguration. See how this works?

@tommysantos14 Anyone who thinks the president drives the economy and that economies are somehow “inherited” needs to go get a better education in economics. The president has very little control of the economy.

💵 Baselined: The Biden Economy We are taking a snapshot of Joe Biden's economy before he leaves office so we have a point of comparison when Donald Trump begins to take credit for it, like he did the economy he inherited form Barack Obama.

@sriramk Long ago and before they apparently went Woke Wired just deteriorated into something not worth reading. It’s not a woke problem - it’s the fact that media has moved to clicks/attention and the quality has deteriorated across the board.

@martin_casado I would also speculate that there is an information theory Nobel prize related to LLMs out there somewhere. There must be some sort of informational limit bounded by something??? Have you seen anyone working on this?

Using the Shannon entropy to guide LLM planning on inference makes so much sense. Really exciting to see the work move in this direction.

@martin_casado I had been thinking exactly this. I wondered whether Shannon information theory applied so this post just partially made my intellectual day!

@vkhosla Economic policy? The rest I agree with. Neither side has an economic policy that makes sense to me.

I may be irrational about Trump but it is hard for me to understand why half of Americans are ignoring the complete lack of values  for a position that should be a role model. And RFK trumpets himself as a "climate warrior" joining "drill baby drill" and ignoring climate, just to get an appointment. I have always disliked him for his anti-science vaccine stance but this "deal making"  (which is illegal if bartered) with no principles is really bothering me. And RFK didn't care if he "did a deal" with Harris (reportedly tried that) or Trump. I hope that is not "deranged" for calling a spade a spade? So again, my priorities 1. Better (not perfect but not despicable values) values 2. climate 3. economic policy.  Then many others. Incidentally, if you think about it,  climate debt (potentially trillions) is hard to pay off. Economic debt can be inflated away in the worst case, not that I am a fan of that.

Let’s talk about what is true: - you’ve cherry picked two Canadian academics to back the bill who are not accountable to it, and not representative of consensus view - you’ve aligned with a fringe group who is primarily concerned with science fiction AI risks - you’ve met with an summarily ignored multiple California academics, founders, and investors And the net is you’re going to hurt California research, and innovation. And you’re giving a hand out to big tech.

@martin_casado You are missing “I am a strange loop” by Hofstadter

Short overviews of writings on consciousness: What is it Like to be a Bat? : Nagle - qualia is a thing The Conscious Mind : Chalmers - science won't explain consciousness Consciousness Explained : Dennett - consciousness is not a thing, or a natural brain function. Shadows of the Mind : Penrose - brain isn't a Turing machine. also quantum The Quest for Consciousness : Koch - consciousness is a property of complexity What books am I missing? I'd love to read more.

@RBReich You are conflating a lack of corporate regulation and inflation. Regulation might help but doubtful that it is the primary source of the inflation problem. And BTW - corporate profits ALMOST ALWAYS go up in absolute terms with inflation but they have rarely been the driver.

The next time you hear someone blame inflation on stimulus checks that Americans got four years ago or workers finally getting much-needed pay bumps, kindly remind them that corporate profits are at a 70-year-high. Hello?

@Dan_Jeffries1 Fantastic post. Well said.

If you actually want to regulate AI, first you need to have a clear, grounded, practical understanding of AI. Then you need to understand the limits of what legislation can and can't do. What legislation can't do is guarantee that something is always safe or we wouldn't have the Boeing MAX incident in one of the most heavily regulated industries on Earth. Those smart, savvy regulatory bodies would have swooped in and stopped the problem before it even happened like the pre-crime unit in Minority Report! On the flipside, heavily regulated industries are slow moving, bloated and have little to no competition. The same will happen to AI if people manage to ram through the kinds of overreaching and misguided bills we're seeing. Too many folks have a black mirror funhouse view of AI. The inevitable outcome of those views will be terrible legislation that crushes innovation and competition while solving exactly no problems. If you start from the wrong premise, everything you do after that it is inevitably wrong. Here's the key to understanding AI and how to regulate it with sound, sane, light touch regulation: - AI is not dangerous in the way Doomers think about it, i.e growing sentient, recursively self improving and treating us like the bugs in 3 Body Problem, aka "humans are bugs, must destroy." It does have some potential dangers in other ways that are barely talked about and those get lost in bombastic rhetoric. - AI is dangerous the way a kitchen knife is dangerous. 99% of the people will cut vegetables with it. 1% will stab someone. We should let most people cut vegetables and punish the 1% of people who stab someone. We do not take the kitchen knife off the market because it can be misused. Everything can be misused. - If you are building AI in a field where people could get physically hurt or killed, such as the medical field or self-driving cars, you should be held to a higher standard of safety. - The nuclear analogy to AI is specious. Nuclear weapons are designed to do one thing only, kill many people at once. AI has a huge range of capabilities and uses. This analogy is absurd. - Attempt to control AI will only lead to a global imbalance where some societies shoot themselves in the foot with overreaching and difficult to comply with regulation, while other countries see an opening and race ahead. This is how empires fall and new empires rise. We will never, ever come to a consensus on AI across the world. - The best regulations usually come after we see what problems happen in reality. As Mike Tyson said "everyone has a plan until you get punched in the face." We can foresee some challenges but mostly we can't see anything meaningful until something has contact with reality. We have a long history of safety bills that were very effective, such as in mining and worker safety, that came after we saw the patterns of problems and moved to address them. - The real dangers of AI are using it for mass surveillance and the military and there are 100% exceptions carved out in every law and proposed bill for military and surveillance use cases. This should shock no one. - Moral panics are delusional and they have a way of escalating and their inevitable outcome is bad/misguided/dangerous legislation or people getting hurt/killed (i.e, witches get burned (actually hung or drowned, for the nitpickers), real people go to jail like in the 80's Satanic panic where a couple went to jail for 28 years with no evidence of child abuse because "Satan made the evidence disappear" (for real). - Irrational fear is a terrible, terrible way to push for policies. Fear has a way of getting out of control and getting away from you. It sweeps out like a storm and it warps and changes as it goes, getting picked up and amplified by other unhinged people, who twist it into their own dark agenda and continue to escalate. What you get is just bad things at the end of that horrible cycle: again, people hurt or killed (and not by AI but by other people) and terrible policies. - Even worse, this irrational AI fear is already being picked up by far right and far left ideologues and being warped into whatever bizarre policy they want to push. The policies they get through with this kind of rhetoric, will have nothing to do with what the Doomers want, which is international cooperation and coordination. It will be used as lip service to justify bigger escalation with China, or sanctions, or anti-immigrant rhetoric. We've had a long history of absurd moral panics for everything from bicycles, to radio, to TV, to video games. They lead to nothing good. - "This time is different." It's not. - Apocalypse predictions have a 100% failure rate in the history of prediction. Predicting AI apocalypse is no different than predicting apocalypse because "God is mad at us and destroyed the crops." It has no validity or evidence to support it, expect in the mind of the people imagining it. Eventually someone may be right by sheer chance but nobody will be around to congratulate them. Much of the rhetoric around AI control and danger is drummed up by people with zero real, practical experience in AI, people who masquerade as researchers in the way the People's Democratic Republic of Korea is democratic and republic. These people are the useful idiots, aka the Baptists to the Bootleggers. - The Bootleggers are the big AI companies who are talking out of both sides of their mouth, claiming AI is dangerous (while they race to build bigger and better ones) and they're using the useful idiots to try to create monopolies and push for anti-business, anti-American licensing regimes to train and run big models that are in nobody's interest but their own. Their actions remind me of Pan Am in the 1950s trying to get a monopoly on international travel. See the new "protect OpenAI from competition bill in California (SB 1047) as the clearest possible example." - Open source/open weights are essential. If you oppose open source, you don't understand the history of the last 20 years of software development where open source is in 95% of software. Linux runs ever major cloud and super computer, your router, most phones, and control systems. It also runs botnets and enemy supercomputers. The benefits of having a common set of building blocks still far, far, far outweigh downsides. You can see more detailed arguments here: x.com/dan_jeffries1/… x.com/dan_jeffries1/… x.com/dan_jeffries1/… x.com/dan_jeffries1/… x.com/dan_jeffries1/… x.com/dan_jeffries1/…

@gruber I thought this was what he has been doing?

If Trump were smart he’d make an Apprentice-style game show out of this. Make the VP candidates demean themselves performing stunts and kissing his feet on camera every week. Then at the end of each episode Trump fires one of them, tells them why he doesn’t like them, and they walk-of-shame out of Mar-a-Lago into an Uber X. The coda is backseat camera footage of the loser telling the camera how much they still love Trump. politicalwire.com/2024/04/12/tru…

@martin_casado This is so true and frustrating. An incredible technology appropriated by too many snake oil salespeople.

OH : "95% of what we do in AI is very simple, but the literature way overcomplicates it by using mathematical-looking nonsense whose notation is all over the place. In general it is beset by unclear thinking and trying absolutely random things."