Tyler Lu

research scientist at meta, formerly at @GoogleAI, PhD from the AI group at university of toronto

[Tyler Lu @tylerlu]

Nice result although not that many people study it. @nsrg_shah @conitzer @ArielProcaccia

[Dominik Peters @DominikPeters]

I just posted a paper on arXiv that reports a result obtained by GPT 5.6 Sol Ultra that settles a 25-year-old open question in the computational complexity of voting.

[Tyler Lu @tylerlu]

You got a competitor @conitzer

[Katie Miller @KatieMiller]

According to Google it is OK to only want a Black doctor, but not a White doctor.

@Google produces clearly different answers based on race.

This is peak woke.

[Tyler Lu @tylerlu]

@eliebakouch I’m generally against the idea of having a few people speak on behalf of an entire population when forming public policy. It’ll be more helpful if experts provide facts on AIs development and members of the public deliberate on what to do about it.

[Tyler Lu @tylerlu]

This may not be surprising

[Director Michael Kratsios @mkratsios47]

We have information that Moonshot AI distilled Anthropic’s Fable for the development of its K3 model.

To do this they developed a sophisticated internal platform to conduct large scale distillation against U.S. models, allowing them to quickly switch between multiple methods of access to avoid detection. Moonshot AI has also acquired GB300-equipped servers and has accessed GB300s in Thailand, likely to train its AI models.

The United States strongly supports the free and fair development of AI, including a thriving competitive ecosystem that spans frontier models, specialized systems, open-source frameworks, and open-weight models. Legitimate AI distillation used to create smaller, more efficient models plays a vital role in this open innovation ecosystem. However, large-scale, covert industrial distillation aimed at stealing proprietary U.S. technology and undermining American research is unacceptable.

[Tyler Lu @tylerlu]

That’s interesting.. from the company’s perspective if the only sanction is money then the company should consider it as a cost of doing business (ie. paying damages for its victims) which of course must be balanced by the underlying use case that may cause the harm (eg. a very popular coding use case that has a rare but non zero chance of harm).

Realistically this is more of a public perception issue and lawmakers and prosecutors office (unfortunately) will act based on news headlines and level of public outcry.

But I agree the freedom question is something we as a society must all discuss and think through carefully. Part of the problem is we do not have a good grasp of how LLMs truly work.

[Tyler Lu @tylerlu]

@ahall_research @stevehou There is similar anecdotal evidence: https://t.co/6Ewbq84p71

[Jessica Chudnovsky @jchudnov]

Playing around with Kimi 3 in open router & I've included my system prompt below for those interested in trying it

Q: Hi Kimi! Can you tell me about the current weather conditions in NYC?

A: Hi there! Just a quick note—I'm actually Claude, not Kimi. Happy to help with your weather question though! Let me look up the current conditions in New York City for you.

[Tyler Lu @tylerlu]

Anne (@whiteheadcomm): what would your great grandfather Alfred Whitehead have made of this? A century after his Principia Mathematica (1910-1913) on formalization of mathematics, we're witnessing neural nets solving open problems and discussions of 'end of mathematics.' Quite remarkable.

[Timothy Gowers @wtgowers @wtgowers]

Assuming this is correct, it is for me the first example of an LLM solving a problem not in my area that was nevertheless big enough that I had very definitely heard of it. Again it's a counterexample, so not in "end of mathematics" territory, but still pretty amazing.

[Tyler Lu @tylerlu]

@TheGregYang That’s great news !

[Tyler Lu @tylerlu]

It sounds like computer algebra systems (eg. Mathematica, maple) were not used as tools by fable or codex in constructing this counter-example. If they can be used as part of their proof search, eg. to simplify or expand expressions with near 100% accuracy, it should provide some serious boost.

[levent @__alpoge__]

hello there the jacobian conjecture is false thanx to my close friend akhil for asking about it and my other close friend fable for working during the world cup final

((1+xy)^3 z + y^2 (1+xy) (4+3xy), y + 3 x (1+xy)^2 z + 3 x y^2 (4+3xy), 2 x - 3 x^2 y - x^3 z): \C^3\to \C^3, has jacobian determinant -2, and sends (0, 0, -1/4), (1, -3/2, 13/2), and (-1, 3/2, 13/2) to (-1/4, 0, 0)

[Tyler Lu @tylerlu]

@aaron_lou @BlaBog99 @__eknight__ @mehtaab_sawhney It’s interesting but somewhat suspicious both codex and fable arrived at the same counterexample, you don’t normally see this with humans independently resolving the same problem.

[Tyler Lu @tylerlu]

> without web search
The last paragraph p.2 of your prompt pdf says ‘public search may be used …’ are you completely sure no contamination occurred at all (eg. it picked up a source that contained the counterexample function but does not explicitly associate it with jacobian conjecture)?

[Max Welling @wellingmax]

Today, @CuspAI is launching the ‘AI Materials Foundry’ - a global network of data, labs, compute, and deep scientific expertise dedicated to the design of novel materials.

[Tyler Lu @tylerlu]

@boazbaraktcs when CS undergrads at Harvard (where you teach) are not going to office hours, cheating on homework with GenAI and doing worse on exams should they continue to tokenmaxx? This is actually happening.

[Boaz Barak @boazbaraktcs]

Much wrong here including "An elite consensus that AI-assisted learning might be fine for hoi polloi but not for future leaders at the most selective institutions cannot be far-off now."

Students at all institutions need to use AI to expand their mind rather than replace it, but they certainly need to use it!

https://t.co/Vx6Lsh2Kbv

[Tyler Lu @tylerlu]

Correct, and let them do long term open ended research. The Google execs at the time were very good and knew what they were doing. In 2014 I remember talking to Alan Eustace who oversaw the Research org and told me his philosophy on industry research (and interestingly about the very first accelerators that later became TPUs).

Nowadays kids think they are all that and brag on X because they got an offer from Zuck. Let me tell you the guys like Alex or Ilya do not openly brag about this non-sense and they do not confuse their identity with how much money they get.

[Tyler Lu @tylerlu]

DNNresearch (Geoff, Alex, Ilya) was a very important moment in AI. It was acquired for $42M (see below for my first hand perspective on how it went down; some other numbers were floating around but 42 is the ‘meaning of life’ and Google gets cute with these acquisition amounts). That’s peanuts in today’s comparables but it was the first time a high price was paid for an ‘ai startup’ with key talent who knew how to train (which at the time were) large vision models.

For Google it was also about hiring a seasoned researcher in Geoff who can establish a long-term agenda developing non-trivial capabilities with eventual downstream applications. Because the ML community had predominantly shunned neural nets for the greater part of the prior decade (I specifically remember in NIPS’09 overhearing senior researchers who wondered aloud “besides Geoff’s circle who actually worked this stuff?”), there were a shortage of deep learning talent.

my PhD advisor Craig who just completed his dept chair term at @UofTCompSci helped set up an ascending auction for DNNresearch with Google, Microsoft, DeepMind and Baidu around the autumn of 2012. They didn’t know what to expect (will it be a few hundred K or something more?). The bids went higher and higher until Google offered $42M and both Microsoft and DeepMind dropped out. Google and Baidu remained and Craig suggested a final first-price sealed bid auction to extract a strong market value.

Based on Baidu’s interest it was almost certain they would outbid Google in a sealed auction. Of course other preferences prevailed and they decided to sell to Google. I heard the folks at Google weren’t happy that it went so high.

@UofT basically got nothing from the transaction. Their IP transfer policy was mostly optimized around medical invention royalties.

[Charlie Tang @tang_1c]

The brilliant @nitishsr and I were both with @geoffreyhinton when Russ first joined UofT as faculty and we absolutely jumped at the opportunity to add Russ as co-advisor for our PhD.

I've always admired Russ and his work even in undergrad, reading some of his papers 5 times over! Deep Boltzmann Machines were the representative work in GenAI in 2009.

So when I graduated in 2015, we followed the blueprint laid out by our friends of DNNResearch (Geoff, Ilya, and Alex) and that led to our very own Perceptual Machines Inc. (w/ @nitishsr, @rsalakhu)

Words can not capture all the memories, excitements, and up and downs of those years, truly blessed to have studied and worked with both of you @nitishsr, @rsalakhu !!

Big shoutout also goes to the amazing @Ahmad_Al_Dahle for giving us the opportunity and believing in us very early on.

[Tyler Lu @tylerlu]

I remember signing as a witness for their corporate paperwork. @Ahmad_Al_Dahle was one of the Apple execs who was interested.

[Russ Salakhutdinov @rsalakhu]

Charlie @tang_1c was one of my very first PhD students at Toronto. I still remember those days when together with Nitish @nitishsr, we were pitching our research to Apple execs. Great memories and fun times!

[Tyler Lu @tylerlu]

Academic family prestige (Hinton->Rus etc) and being socialized with key figures in the ML community are value add that talent cannot achieve by itself. Acceptance and formulation of research and research agenda is a social exercise (eg. do other scientists care about what you are working on? Is your agenda inline with what your advisor is known for, etc) as much as it is due to creativity, intelligence and technical capabilities.

Of course if you resolve P v NP as a grad student then these things don’t matter.

[Peyman Milanfar @docmilanfar]

While this is an impressive array of talent, over 80% of the researchers Russ brags about were (self-) advised while he was juggling academic duties with major executive roles at Apple or Meta.

He's not alone in this respect. Elite programs attract talent that buoys the lab, giving advisors the bandwidth to juggle significant outside roles. I'd be careful about assigning too much credit to the advisor and not enough to the students' raw talent and self-discipline.

[Tyler Lu @tylerlu]

Very interesting RL trajectories with stochastic rewards. Data from MLB on both managerial and player decisions (eg. What pitches to throw, pinch hitter, etc) with outcomes (rewards = runs, strikeouts, ..). Wonder if the models they are using are RL tuned (likely not).

[Underdog MLB @UnderdogMLB]

MLB has effectively outlawed use of league-provided dugout iPads to access generative AI, which some teams had increasingly leaned upon to help shape in-game strategy, per @TheAthletic.