andyxandersen.bsky.social @andyxandersen.bsky.social · 28/12/2024We need criticism, sure. This is a debate about rate of progress and technical choices. That's where the focus should be. Don't need heavy metaphors, don't help. 000
andyxandersen.bsky.social @andyxandersen.bsky.social · 28/12/2024No need to make the contrast so stark. The hype is high, but the improvements are also steady and more than what some skeptics acknowledge. 020
andyxandersen.bsky.social @andyxandersen.bsky.social · 28/12/2024The world is too complicated to get fully general and fully correct systems from the outset. For decades we had very narrow systems that failed as soon as something changed a bit. Now the systems are very general, but will need a lot of refinement. 000
andyxandersen.bsky.social @andyxandersen.bsky.social · 26/12/2024AI is becoming less efficient, short-term, because the problems it is trying to solve are becoming harder. The current approaches are not optimal, but a tried-and-true solution is to succeed first and reimplement later. 110
andyxandersen.bsky.social @andyxandersen.bsky.social · 21/12/2024True, but this is a first-cut approach. Longer term, AI will learn to classify the problems better and be more efficient in the approaches. Then, unlike pure LLM, but like humans, harder problems have to take longer. 120
andyxandersen.bsky.social @andyxandersen.bsky.social · 20/12/2024Francois Chollet states, and I agree, that for hard problems that were not seen before, there is no magic, neither for people, nor AI. One must search around. So, what we see now is a big deal. That some of the test is private is also important. 000
andyxandersen.bsky.social @andyxandersen.bsky.social · 20/12/2024I read Melanie Mitchell's article before, and I read it this time again. It is important to note that while o3 is not AGI, of course, it is a very important step for doing what people do, which is not recite answers, but actually try hard various ideas till it works. 100
andyxandersen.bsky.social @andyxandersen.bsky.social · 20/12/2024Of course AGI is not nigh. Not clear what "semi-private" means. But this is a solid enough results that can't be dismissed by saying they just subvert benchmarks. 100
andyxandersen.bsky.social @andyxandersen.bsky.social · 20/12/2024o3 scored 87.5% on the Semi-Private Evaluation set of ARC AI. That is a hard benchmark, and its creator, François Chollet is a very serious researcher. 200
andyxandersen.bsky.social @andyxandersen.bsky.social · 20/12/2024Learning to interact with the real world will take a lot of real-time data and better architecture. The current approach is an approximation, but a good one. It interacts with the world in training batches. 010
andyxandersen.bsky.social @andyxandersen.bsky.social · 20/12/2024It is a bad prediction. If you choose to see the cup half-empty each time, that's your choice. The algorithms are getting better as measured on hard benchmarks, such as ARC-AGI. 110
andyxandersen.bsky.social @andyxandersen.bsky.social · 20/12/2024François Chollet is impressed with o3, and has results for ARC-AGI to prove it (on Twitter). The work is always incremental, but the rate is very good. More to do. 020
andyxandersen.bsky.social @andyxandersen.bsky.social · 08/12/2024Casey Newton's piece was bad criticism of AI skepticism. Picking a fight with a person, rather than addressing the arguments of the skeptics, wasn't great. "phony", "sucks", etc, didn't help either. But the industry is in a better place than skeptics like to think. 000
andyxandersen.bsky.social @andyxandersen.bsky.social · 01/12/2024There are many things in the world, however, that are too diffuse and poorly-stated to be baked into software. That's where large-scale neural nets come in. Not to replace software, but to complement it. 000
andyxandersen.bsky.social @andyxandersen.bsky.social · 01/12/2024LLM and RAG are very coarse approximations and lack understanding. They are however a prerequisite. The world is too complicated to only do efficient model-based logic. Need a combination of models and context/cataloging, which LLM offers. (@andnig.bsky.social) 000
andyxandersen.bsky.social @andyxandersen.bsky.social · 29/11/2024The focus now is on in-distribution learning by imitation. That is hard enough as it is and the advances have been good. Spatial intelligence needs more architecture, without necessarily throwing out what we learned so far. 000
andyxandersen.bsky.social @andyxandersen.bsky.social · 27/11/2024To complement that, Ben Goertzel wrote an essay that beautifully summarizes just how hard it has been to make principled approaches to ontology work. Lots of hard work ahead. substack.com/home/post/p-... 000
andyxandersen.bsky.social @andyxandersen.bsky.social · 26/11/2024Here's a very insightful essay into limitations of reasoning in LLM. Its point is that open-ended problems with no model or verification will be very hard to tackle within the existing framework. aidanmclaughlin.notion.site/reasoners-pr... 100
andyxandersen.bsky.social @andyxandersen.bsky.social · 25/11/2024I agree, however, when it comes to "hybrid data-driven/model-based approach to agent design". Agents need to be guided by models, as data alone only provides high-level predictions that need validation and refinement. 010
andyxandersen.bsky.social @andyxandersen.bsky.social · 25/11/2024The bitter lesson remains that the world is too complicated to spell it out. The rate of progress for current data-based methods is very good, though there is a lot more to do. Methods based on principled reasoning made no progress for 40 years. 100
Reposted by @andyxandersen.bsky.socialThe Onion @theonion.com · 19/11/2024Trump Locks Bathroom Door So Elon Musk Can’t Follow Him In theonion.com/trump-locks-... 553291274263
andyxandersen.bsky.social @andyxandersen.bsky.social · 23/11/2024"I stated that point and many people dismissed me. Nobody in the tech elite backed me up." People like François Chollet, Yejin Choi, Margaret Mitchell, Sasha Luccioni, Fei-Fei Li, Demis Hassabis, have been very skeptical of LLM. Surely your voice has a value. But asking for a citation? 030
andyxandersen.bsky.social @andyxandersen.bsky.social · 23/11/2024Chatbots are getting more reliable and more useful. They really help in daily work, and results are good. Soon they will function as agents and robots, which will give them more chances to get feedback and get better. It is steady evolution. 000
andyxandersen.bsky.social @andyxandersen.bsky.social · 21/11/2024The pendulum swung so much left in the last 15 years. It appeared that the recent gains became settled law, and now there's a hard right push. I don't think the upcoming admin will change things as much as they hope though. 000
andyxandersen.bsky.social @andyxandersen.bsky.social · 19/11/2024Seems like in we are in for a repeat of 1st term, with chaos and incompetency, and the dude being his own worst enemy. 000
andyxandersen.bsky.social @andyxandersen.bsky.social · 17/11/2024Demis says: 10 years ago we thought we needed to teach systems abstractions. But apparently, with enough data, systems seem to learn and generalize. We did not expect that would work as well as it did. Even though they don't have a proper model of the world. 010
andyxandersen.bsky.social @andyxandersen.bsky.social · 17/11/2024We need new innovation, yes. That said, the strategy of OpenAI and Google has been very sound. Push scale and LLM as much as one can, and prepare for how to fill in holes in it. Demis has been talking about this for a while. 110
andyxandersen.bsky.social @andyxandersen.bsky.social · 09/11/2024Biden screwed up in that debate royally bad. Forcing him out was the best thing the Dems did, even if it did not help, eventually. Trump is just erratic. He may still become demented but the bjob is not a sign of that. 010
andyxandersen.bsky.social @andyxandersen.bsky.social · 15/01/2024Marcus is a smart man. Recently he's been trolling too much though. He also accused LeCun of being one of the top 5 dangerous people in tech, then got each other blocked. Very juvenile. 020
andyxandersen.bsky.social @andyxandersen.bsky.social · 08/12/2023To have due diligence on data is a fair point, especially if later a company is sued into oblivion. Historically, however, the parties usually settled quietly if a violation was found. Many of the claims by artists will likely not stand up in court. For a company moving fast can be an advantage. 000
andyxandersen.bsky.social @andyxandersen.bsky.social · 08/12/2023The "horse" is the profit a company makes. That's what pulls the "cart". Social responsibility is nice, but let us face it, the industry would not exist without a profit. That need not be "alienating". 100
