Gary Marcus

Three thoughts on what really matters: 1. Fuck cancer 2. Friends are irreplaceable 3. The new "Marcus test" for AI is when AI makes a significant dent on cancer May that happen sooner, much sooner, rather than later. In memory of my childhood friend Paul.

“one 80% price cut ends OpenAI and Anthropic” possibly overstated but directionally correct.

"AI does not have to be built this way" We built the wrong AI. This should be mandatory reading:

@keyonV @UCBStatistics @Berkeley_EECS congrats! keep up the great work!

Thrilled to share I'll be joining UC Berkeley next year as an assistant professor in @UCBStatistics and affiliated with @Berkeley_EECS. My research will build methods to test/improve the implicit world models of AI systems, so that they reflect reality and human understanding.

vile if true

This AI recursive self improvement (RSI) paper shows no sign of fast takeoff. The AI model advances at about [Intelligence]^0.075, or the the 13th root of input intelligence [1,2]. That means the intelligence explosion fizzles out rapidly. 1 - I expect this team or others to do much better than this over time! But no sign they'll get to an exponent even approaching 1.0, which is what you need for runaway self-improvement. 2 - The actual rate of self improvement is probably even worse than the math shows, as this is improvement on a specific metric rooted in a specific training set which taps out at a finite capability cap. Actual improvement on some training-set-independent scale of intelligence is probably log-linear, not merely power-law scaled. I'm super impressed with this work and the team, to be clear! Just not seeing a Singularity on the horizon.

“I asked a few AI researchers whether they could name any other real-world software that scales so poorly. None of them could think of any. Even outside the world of software, it’s hard to find a comparable example, given that economy of scale is the principle that has made light bulbs, cars, and clothing so affordable. By economic and engineering measures, generative AI might be the worst technology ever deployed” @_alexreisner @TheAtlantic In “Generative AI is an engineering disaster” theatlantic.com/technology/202…

I want to have this in writing so I can refer back to this tweet before it inevitably becomes a consensus view on twitter and say 'I told you so!' - The mass majority of researchers and academics in deep learning not only do not understand the statistical nature of the field, but are completely clueless as to where true intelligence will come from. AGI, and indeed any general intelligence will not come from deeplearning for the same reason that being able to extrapolate a 4th point from a set of 3 points you've interpolated is impossible - extrapolation from solely interpolation over a bounded domain is impossible. @GaryMarcus put it best in his 2018 work (which I highly recommend you read here: arxiv.org/pdf/1801.00631), Deep learning fundamentally lacks the framework necessary to learn abstractions. AI as we know it today is at the very core, simple polynomial regression over a known data set that we approximate over. To ever go beyond this, you either a) must continually expand the data set that you interpolate over (which is commonly referred to as the notion of 'continual learning') or b) discover a new paradigm in which abstractions are genuinely learnable and understood by whatever system is encountering them. That being said, AI I argue is now going through a dark age. Much to the same way symbolic AI under the actions of Marvin Minsky sucked all the air out of the room and left connectionism and all other paradigms of AI dead, Deep Learning is now doing the same. All new startups, academics, researchers, etc - are all converging on the exact same approach. Take a deep learning network, and either scale it (like we've seen with the embarassingly bad performance of 'just scale a VLA!' from a number of high profile labs), or try to simply attach a harness around deep learning network and pray it all goes away. Genuine innovation in the field has more or less completely been pushed out by people attempting to make mediocre increments on a dead approach simply because it less scary to do so, and because observing marginally better improvements is more calming than going through the wilderness of trying to discover a real solution that most people are unaware of. Evidence to the above is obvious. By now, we've seen over a dozen startups every quarter come out with ludicrously large rounds with the operating philosophy of 'just throw data' or 'talent' at the problem and known solutions but in time I predict that most of these companies will not only have failed, but likely have produced little of substantive value. The amount of papers in NeurIPS for example that are all essentially making the same claim of 'here's how we hyperparameter tuned to this specific task and observed a 10% increase on this benchmark we gamed over SOTA' has become so high, that finding actual novel work in the field has now become an extremely difficult task. To that end, I'll also close off this short essay by clarifying, I don't think deep-learning is useless. LLMs and indeed transformers, CNNs, LSTMs, and other architectures have been used to do very useful and tangible things, whether its write good code, recognize fraud, etc. I do not mean to insinuate that deep-learning is useless, but the best analogy to give is everyone is scrambling to get flight, but we're all focused on building rocket engines instead of wings. With @RichardSSutton's recent announcement, I felt even more inspired and vindicated in my viewpoint. I've long held this view, and have been called crazy at every step of the way. But now as more people within the field cannot deny the ugly truth we face ourselves with; I increasingly receive more messages from friends whether researchers or founders all saying the same thing 'you were right.' Of course, like most - we (the company we've founded) believe in a particular approach that will work based on a combination of methods, namely around neuro-symbolic approaches and evolutionary algorithms. Only time will tell how right we are. The future of AI will very interesting.

@stephenbklein 💯

@GaryMarcus It will be a mess but the sooner the better because it will just keep getting worse and the sooner it happens the sooner we can get through it and start building tech that helps people rather than just create share and raise capital

little by little, OpenAI’s storytelling is falling apart. my 2023 projection that they would someday be viewed as the WeWork of AI is looking stronger by the day.

multiple such incidents have been reported. a clear reminder that current AI cannot be trusted. in racing these techniques ahead, we are asking for trouble

I work at Google DeepMind. This won't make me popular. But it's all public reporting: 2014: DeepMind reportedly sold to Google on conditions: no military use, independent oversight 2026: a Pentagon contract for "any lawful government purpose" Not one safeguard survived intact

Is Oracle the canary in a coal mine?

$ORCL is down 57% since my September essay “Peak Bubble”:

@Blum_OG fantasy is doing the heavy lifting

@GaryMarcus the satellite part is doing all the heavy lifting in that valuation

what did you expect from a meme stock (with a satellite company thrown in)?

@GaryMarcus The data is confirming what skeptics have been saying since day 1 - the gains are not justified by the cost. The hype reckoning is coming.

Seismic! “current deep learning methods are weak and inefficient, and need not more tweaks, but fundamentally new ideas and a thorough reworking before they can provide a solid foundation for achieving the more ambitious goals of AI” — incredible to see @RichardSutton so forcefully affirm what I have been saying since 2018, in “Deep learning: A Critical Appraisal” an arXiv paper that @ylecun once alleged was “mostly wrong”.

@HassanRIsmail @RichardSSutton @ID_AA_Carmack @kjaved_ at least since 2018, arguably since 1998

@RichardSSutton @ID_AA_Carmack @kjaved_ Deep learning is fundamentally flawed; we've been saying this for almost a year now and have been getting called crazy at every step of the way; so it's refreshing to have you outwardly say it, Rich. I can't imagine how @GaryMarcus must feel if he's been saying it for longer

I can’t say enough good things about John Carmack @ID_AA_Carmack and his Keen Technologies. But now Khurram Javed @kjaved_ and I have broken away to start our own startup and pursue a slightly different path toward understanding intelligence. Like Keen (and like Ineffable) we at Oak Lab @oaklab_ai believe in reinforcement learning and that intelligence is created and maintained from run-time experience. But we think current deep learning methods are weak and inefficient, and need not more tweaks, but fundamentally new ideas and a thorough reworking before they can provide a solid foundation for achieving the more ambitious goals of AI.