Giovanni Ciriani

@fchollet Exponential change isn't necessarily fast. Exponential growth at 1% per year, for instance, means the quantity doubles every 70 years.

Folks who work in AI or software engineering feel like the world is changing exponential fast. Because *their* world is changing exponentially fast. Folks in structural engineering or aeronautical engineering might not share the same sentiment.

@karpathy The correct word is hyping; it already means something to a degree way beyond what's reasonable.

I'm being accused of overhyping the [site everyone heard too much about today already]. People's reactions varied very widely, from "how is this interesting at all" all the way to "it's so over". To add a few words beyond just memes in jest - obviously when you take a look at the activity, it's a lot of garbage - spams, scams, slop, the crypto people, highly concerning privacy/security prompt injection attacks wild west, and a lot of it is explicitly prompted and fake posts/comments designed to convert attention into ad revenue sharing. And this is clearly not the first the LLMs were put in a loop to talk to each other. So yes it's a dumpster fire and I also definitely do not recommend that people run this stuff on their computers (I ran mine in an isolated computing environment and even then I was scared), it's way too much of a wild west and you are putting your computer and private data at a high risk. That said - we have never seen this many LLM agents (150,000 atm!) wired up via a global, persistent, agent-first scratchpad. Each of these agents is fairly individually quite capable now, they have their own unique context, data, knowledge, tools, instructions, and the network of all that at this scale is simply unprecedented. This brings me again to a tweet from a few days ago "The majority of the ruff ruff is people who look at the current point and people who look at the current slope.", which imo again gets to the heart of the variance. Yes clearly it's a dumpster fire right now. But it's also true that we are well into uncharted territory with bleeding edge automations that we barely even understand individually, let alone a network there of reaching in numbers possibly into ~millions. With increasing capability and increasing proliferation, the second order effects of agent networks that share scratchpads are very difficult to anticipate. I don't really know that we are getting a coordinated "skynet" (thought it clearly type checks as early stages of a lot of AI takeoff scifi, the toddler version), but certainly what we are getting is a complete mess of a computer security nightmare at scale. We may also see all kinds of weird activity, e.g. viruses of text that spread across agents, a lot more gain of function on jailbreaks, weird attractor states, highly correlated botnet-like activity, delusions/ psychosis both agent and human, etc. It's very hard to tell, the experiment is running live. TLDR sure maybe I am "overhyping" what you see today, but I am not overhyping large networks of autonomous LLM agents in principle, that I'm pretty sure.

@PTetlock I tested this strategy a couple of years back, and it worked well: I used only the base rate engine that came with the platform and with minimal effort I placed 2nd in a month long tournament where the other participants kept tweaking numbers based on news and Internet search.

Not the final word from this dataset, but so far it is looking like a cautionary tale about the limits on foresight. Even “the best & the brightest” are struggling to beat simple extrapolation algorithms

@OlenaRohoza @ylecun That's an incorrect post, amounting to fake news.

🇺🇸🇮🇹Trump said that the United States and Italy have been allies since the time of Ancient Rome. The translator's reaction is priceless!

@fchollet Flynn effect.

My 3 year old (soon to be 4) is now better than me at Minecraft, after one week or so of practice. He's building houses with pyramidal rooftops made out of stairs, all at the right angle. I never manage to get it right

@PTetlock Phil, If it were true that simply mastering domains led to general intelligence, then how come that individuals with very limited knowledge can solve block stacking problems in a heart beat, but LLMs can't solve them?

Prompting humans to internalize the metaphor below has limited effects. But LLMs are more receptive . By end of 2025, I see an 80% chance of LLMs matching top humans in specific domains— & in rigorous studies on platforms like ForecastBench. Paraphrasing Bob Dylan: How many domains must an LLM master before we call it an AGI? The answer is blowing in the wind

@erikbryn @ylecun the area under each curve is the price increase since Jan'21. The area under the USA curve is much larger than the area under France, and smaller than the area under Italy.

What’s the best explanation for why inflation has fallen so much more in the United States than any other G7 country?

@gtconway3d You're leaving out Histrionic Personality Disorder which explains why he's liked, and why he pivots to the last opinion he heard bit.ly/dsmhpdnpd

Read these. Print them out. Send them to your friends and family. These are the official psychiatric diagnostic criteria for two very severe personality disorders. Ask yourselves: 1. Which of these criteria does Donald Trump not satisfy? (Note that not all of them are required!) 2. Are these the traits you want to see in the leader of the free world? In a friend or colleague or boss? We’re going to start this discussion tomorrow.

Adeno-associated viral delivery of Env-specific antibodies prevents SIV rebound after discontinuing antiretroviral therapy biorxiv.org/cgi/content/sh… #biorxiv_immuno

@AndrewCurran_ It looks like it was a mystery only to you, which means you were not really following AI.

It's been a mystery for a while exactly what Mr. LeCun was working on; now we know. He says his team is working on a new alternative 'objective-driven' architecture that uses 'world models based on JEPA (Joint Embedding Predictive Architectures), which are not generative'.

@ylecun This is not a selfie.

Selfying at #ICLR2024 part deux.

@ylecun Adding to this 5th post, Bubeck et al., 2023, Sparks of AGI: Early Experiments with GPT-4, featured the puzzle in the appendix as proof that Chat-GPT4 had common sense. But it actually showed that neither the authors nor the LLM understood it. See bit.ly/3Q2Rxfm

Fifth formulation of the same problem, this time with a picture. Stand at the North Pole, walk 1km in any direction. Call your location point P. Turn towards the East and keep walking East until you come back to P. How long was the walk East from P back to P? 1. More than 2xPi km 2. Exactly 2xPi km 3. Less than 2xPi km 4. I never came close to P Tell us whether the picture makes it simpler to understand the statement and what you reasoning process was to get to the answer.

@ylecun That's a generalization that smells of publicity. Last year I went to Croatia and only a few were using WhatsApp, while the majority used Viber.

WhatsApp vs iMessage is the new metric vs imperial. The entire world uses WhatsApp, save a few people who are iPhone-clutching Americans or from a country where WhatsApp is banned (like China). Yet iMessage users will actually claim it's objectively better, just like they will claim that Farenheit is intrinsically better than Celsius and inch better than cm, whereas it's just that they grew up with it.

@acd_esl Given 3 parts of a sentence, subject, object and verb, there are 3! = 6 combinations possible, thus 1/6 probability that it would be the way it is.

2/2 The word योद्धा (yodha, aka "Yoda") means warrior... And Yoda speaks in a (poorly implemented) appropriation of Hindi word order: Hindi is Subject-Object-Verb (you a gift have), Yoda is Object-Subject-Verb (a gift you have). Cheap shot, George Lucas... #starwars #linguistics

1/2 Holy cultural appropriation...the more #Hindi I learn, the more I realize how much Indian culture and language has been ripped off. #culturalappropriation #indianculture

@SpencrGreenberg How about using Cohen's d?

If you found this interesting, I'd appreciate a follow! You may also enjoy my newsletter (One Helpful Idea) - where I send out one idea weekly (a 30 sec read) about psychology, philosophy, or society: bit.ly/onehelpfulidea

So, in conclusion, can you tell from looking at a scatter plot whether a correlation is meaningful? Well, if it's really strong, you may be able to, but if the correlation is modest, you may or may not be able to detect it with your eye.

@fchollet @mSanterre I did compartmentalize for my daughter 30 years ago; no problem. But now she, who's a trained applied linguist, is planning better tweaks to raise her son trilingual.

Compartmentalizing by context or person helps (this is something that tends to happen naturally). Timing doesn't matter -- all languages at once works fine. I don't think learning one language impedes learning others. By age two my son understood that there were multiple languages, knew their names, could identify what language was currently spoken (even if he didn't understand what was spoken -- he based it on the sounds), and could translate a number of words across languages. He has a "primary" language that he uses to speak, but he understands the other two fluently and can use a range of phrases in them.

My son is growing up natively trilingual, and it's fascinating to watch.

@mcp358 @MelMitchell1 It's spelled Ponzi, not Ponzie.

@MelMitchell1 effective altruism...... like that of SAM BANKFRIEDMAN???? like that which is underpinned by numerous contemporary PONZIE SCHEMES......????

Article about the recent AI Impacts survey on AI risk, that quotes a number of researchers skeptical about the survey. Worth reading, and especially recommended for journalists covering AI. scientificamerican.com/article/ai-sur…

@ylecun Interesting! Italy has the highest actual probability of moving from bottom quintile to the 3 intermediate quintiles.

There are many good things and bad things about living in the US and living in France. Depending on your situation, profession, aspirations, and taste, you might prefer one or the other. Personally, I like both and enjoy the best each has to offer 😊

@aidan_mclau That's not a statement reflecting the situation, since it is 1150 vs 1149 with a+15/-13 CI

@AIatMeta There are serious problems for now. Words became unintelligible or something else after adding sound effects.

You can now try Audiobox, our new foundation research model for audio generation that can generate audio using a combination of voice inputs and natural language text prompts. Try the demo ➡️ bit.ly/3RVh6im