Vivek

@rronak_ arxiv.org/abs/2601.19897

My favorite paper of 2026 so far 🔥 They took On-Policy Distillation (ie the Thinking Machines blog post), but then showed that the policy can be both the teacher and the student model. The idea is to condition the teacher off of a golden trajectory, and then train on the conditioned logprobs of the same model. The crazy part is, you can literally condition the teacher on anything!! This opens up an entire pandora's box of bridging prompt optimization/ICL + weight optimization that I'm very excited about for continual learning Authors: @IdanShenfeld @MehulDamani2 Jonas Hübotter @pulkitology

@PeterBerezinBCA Barrels! 🤣🤣🤣 If it's "under condition of anonymity" it must be true.

Wait, they think that oil is actually shipped in barrels? The stagecoach is ready…

@RealJimChanos Devil take the hindmost.

@ivanburazin your rite

If you’re taking a break for the holidays your ngmi.

@CliffordAsness Newsflash: RFK Jr. claims breathing air causes cancer. Anyone who's had cancer has breathed air in the past year. So... 🙃

From the article: “There are still no studies that support the claim that any of the 20 doses of the seven infant vaccines recommended for American children before the first year of life do not cause autism.” Not serious people in charge of serious stuff. “When did the vaccines stop beating their spouses?”

@Dan_Jeffries1 @akshay_pachaar This logic fails under the fact that that paper wasn't earth shattering at the time it was published in NIPS. Significant, yes. Earth shattering, no.

@akshay_pachaar Cool paper but the reason we know this is not Attention is All You Need v2 is because Google will never ever ever ever ever let their researchers publish something as broad and impactful as AIAYN ever again. :) They only let them publish lesser papers now.

Google just dropped "Attention is all you need (V2)" This paper could solve AI's biggest problem: Catastrophic forgetting. When AI models learn something new, they tend to forget what they previously learned. Humans don't work this way, and now Google Research has a solution. Nested Learning. This is a new machine learning paradigm that treats models as a system of interconnected optimization problems running at different speeds - just like how our brain processes information. Here's why this matters: LLMs don't learn from experiences; they remain limited to what they learned during training. They can't learn or improve over time without losing previous knowledge. Nested Learning changes this by viewing the model's architecture and training algorithm as the same thing - just different "levels" of optimization. The paper introduces Hope, a proof-of-concept architecture that demonstrates this approach: ↳ Hope outperforms modern recurrent models on language modeling tasks ↳ It handles long-context memory better than state-of-the-art models ↳ It achieves this through "continuum memory systems" that update at different frequencies This is similar to how our brain manages short-term and long-term memory simultaneously. We might finally be closing the gap between AI and the human brain's ability to continually learn. I've shared link to the paper in the next tweet!

@Abude_al_asaad On this note, let's change the personal finance rule: invest 100 + your age in equities funded by leverage. Wouldn't that be great?

Our first post about Basic Capital reached over 25M people. The conversations that followed made one thing clear: the way Americans save for retirement is fundamentally broken. (1/

@TMTMoats Soooo..."this time it's different" argument or "we're not there yet" argument? Likely the latter.

Eric Sheridan (GS Internet Analyst), thoughtful response when asked if we are in an AI Bubble: "Do I see signs that point me back to the late 90s or the 2007 time period? Sure. Private market valuations are well ahead of public market valuations. Public market valuations are above historical norms, but they're also below the peak of public market valuations that you saw in 1999 and 2000. Capital market activity is still well below the levels that were seen in 2020/2021, 2007/2008, & 1998/1999. I would argue there are signs of exuberance - there are signs that rhyme with past periods of time, but I wouldn't necessarily align it perfectly with the lessons we've learned in prior periods, at least not yet. That's arguably a duration narrative, that we're just not there yet - that is one framing I would give. The other thing that is a little bit different than prior periods is that typically the companies that generate no profits, and in 1999, it was companies that generated no revenue, are the ones that were driving the most exuberant valuations in the market. The Mag 7, most of those companies, generate free cash flow, buy back their stock, and pay dividends. There were very few companies buying back their stocks and paying dividends in 1999, so there are some key distinctions I would keep in mind as opposed to just drawing a straight analogy."

@ALEXEIMARTOV Great points. This is the gem: "but it is hard to tell if the data center capacity will be better used with new ones at that point especially considering the cost of power / the power efficiency of the new gpus."

people act like the depreciation on GPUs is really difficult, so i made this simple guide. you're the gpu maker: we've gone from a 2 year replacement cycle to an 18 month one, pump the stock pls. you're a growth ceo talking to investors: those gpus last 12 years, lend us money pls. you're the growth ceo talking to your CFO: depreciate those gpus as fast as possible, fuck paying taxes. you're the investor shorting the stock: those gpus only last 2 years, look at what the CFO just said! you're a mega cap CEO trying to protect your P/E ratio / bond rating, which is more important to you than tax optimization: those GPUs last 8 years you're an investor longing the stock: those gpus last 8 years. the reality: the GPUs probably last a little over 5 years but it is hard to tell if the data center capacity will be better used with new ones at that point especially considering the cost of power / the power efficiency of the new gpus.

@FromValue Doesn't make sense. There doesn't seem to be any correlation with and without any lag.

If you hear someone say the S&P 500 PE ratio is historically high, show them this. Context and knowledge.

@junkbondinvest @verdadcap Aren't PE owned companies more leveraged than S&P or Nasdaq companies? Deleveraging to make it on par with S&P or Nasdaq would make PE performance even worse.

Private equity vs public markets scorecard: PE 1-year: 8.1% S&P 500: 25.0% VC 1-year: 6.2% Nasdaq: 29.4% But hey, at least you can't check the price daily

@curiouswavefn Who in this world refers to themselves by their last name?

Amazing story about how intensely Lincoln studied Euclid's "Elements" to find out the precise meaning of the word 'demonstrate'.

@LOGOinvestor @bgurley Bill very likely gets this. But it gets in the way of his IPO outrage narrative.

@bgurley I doubt. Heard in the boardroom pitch “we price it too low, we only release a small amount of stocks, and it goes vertical…then you can do multiple secondaries and pad your balance sheet”!!

With $FIG we have another massive "POP" highlighting the gross inefficiency in the modern IPO process. It's very simple. They REFUSE to match supply/demand (that happened today). They brag about the mis-match - "30X oversubscribed." The outcome is expected & fully intentional.

@greg_ip Doesn't AI bubble being a confounding factor make this comparison quite a bit less kosher?

One reason tariffs haven't affected the stock market more is that they primarily affect the "old economy". GM has $187B in sales but a market cap of just $50B, while Nvidia has $135B in sales and a market cap of $4.1 trillion. GM is much more exposed to tariffs than Nvidia.

@george__mack That attitude seems to be a sign of insecurity.

If someone asks you “what do you do?” within seconds of meeting you — don’t be annoyed. Smile. You’ve just identified a transactional player. Instead, give them the lowest status answer. Then drop your highest IQ insights for the rest of the conversation. This breaks brains.

@CliffordAsness No volatility launderer left behind!

Nobody should be excluded from having their volatility laundered regardless of race, creed, color, or income. Finally our long national nightmare is over.

@eigensteve Love to proofread sections of it.

New Book & Video Series!!! (late 2025) Optimization Bootcamp: Applications in Machine Learning, Control, and Inverse Problems Comment for a sneak peak to help proofread and I'll DM (proof reading, typos, HW problems, all get acknowledgment in book!)

@davidasinclair But is is causation or correlation? Which type of muscle: type 1 or type 2 or both? Is it the most important factor? Compared to VO2Max?

Muscle mass predicts lifespan better than BMI or weight - so get moving and get lifting

@InvestingVisual How did you reach the conclusion that "AI chip market is set to grow from $96B to $927B by 2034"? Any sources? I'm just curious regarding the veracity of this statement.

The AI chip market is set to grow from $96B to $927B by 2034 — a massive 29% CAGR. Businesses best positioned to benefit: • Tier 1 - AI Foundation: $ASML, $TSM, $AMAT • Tier 2 - AI Chips: $NVDA, $AMD, $AVGO • Tier 3 - AI Infra : $NBIS, $CRWV, $CDNS

@davidasinclair Why 14 hours and not 13? What's the mechanism? Yeah you do the research but where is the research paper that shows this?

Fasting for 14+ hours activates autophagy—your body’s cleanup mechanism