DavesNews

AI is my fascination. Reported broadband and the web around the world since 1999. Sometimes get things right. @alnews_wire @aimed4all [email protected]

[General Catalyst @generalcatalyst]

Meet GC

[DavesNews @DavesNews]

For the past several years, private equity’s annual returns have been lower than the S&P 500’s. NYT

[pdawg @prathamgrv]

I’m halfway through this book and the clarity around training LLMs is unreal. If you’re building or researching with LLMs, this one’s hard to skip.

[Balaji @balajis]

This is the EV flippening.

Electric vehicles now fill up roughly as fast as gas. The new BYD chargers add 400km of range in just five minutes.

[hardmaru @hardmaru]

I doubt that anything resembling genuine AGI is within reach of current AI tools—Terrence Tao

“Perhaps this can be resolved by the realization that while cleverness and intelligence are somewhat correlated traits for humans, they are much more decoupled for AI tools (which are often optimized for cleverness), and viewing the current generation of such tools primarily as a stochastic generator of sometimes clever - and often useful - thoughts and outputs may be a more productive perspective when trying to use them to solve difficult problems.”

[Jen Zhu @jenzhuscott]

Humanoid parts supplies by country. The US has an edge in compute but everything else is pretty much dominated by China.

[DavesNews @DavesNews]

From NBC: Between Sept. 17 & Oct. 17, Musk posted, 1,716 times X account, 55 times a day,
Extreme ideas, like the antisemitic “great replacement” theory, which says there is a plot organized by Jews to replace the whites populations of the US & Europe with nonwhite people.

[Brad Stulberg @BStulberg]

A massive new study on peak performance included 34,000 international top performers: Nobel laureates, renowned classical music composers, Olympic champs, and the world’s best chess players. It shows early specialization is a trap, and the road to greatness is long and varied.

[DavesNews @DavesNews]

"all of your paranoid fantasies are actually based in reality." former CIA agent Peter Matthiessen
https://t.co/xPhlIxLCuo

[Andrej Karpathy @karpathy]

@drfeifei @QEPrize I love that the prize recognizes foundational work in all 3 pillars of AI (data, compute, algorithms), congratulations Fei-Fei!!

[DavesNews @DavesNews]

Matt O'Brien's AP headline "US gathers allies to talk AI safety" on target. Gov lying calling it "Global" when half world is excluded. "US ahead in AI innovation, easily surpassing China" wrong even if Stanford says so. Ping me for more

[DavesNews @DavesNews]

Doug - Haven't seen data centers shutting down in Virginia or elsewhere. No problem building new centers in the US. No policy reason they have to be in N. Virginia so no crisis. Powerco bs for rate increases, new builds, keeping dirty plants. May be low price for businesses

[Doug Mohney @DougonIPComm]

@DavesNews That's funny. Try getting power in the Northern Virginia data center corridor.

[Rohan Paul @rohanpaul_ai]

This paper comes back to show again, that LLMs are just pattern-matching machines, not step-by-step reasoners when solving math problems.

Reveals that LLMs don't actually perform step-by-step reasoning when generating direct answers without showing their work. Through probing experiments on arithmetic problems, researchers found that models skip intermediate calculations and rely on pattern matching, challenging assumptions about implicit reasoning capabilities.

-----

🤔 Original Problem:

While Chain-of-Thought (CoT) prompting improves reasoning but slows down inference, many researchers assumed LLMs could reason implicitly without showing steps. However, implicit reasoning consistently underperforms explicit CoT, raising questions about whether LLMs truly reason step-by-step internally.

-----

🔬 Solution in this Paper:

→ Researchers used Qwen2.5-72B to solve multi-step arithmetic problems without showing work.

→ They probed the model's hidden states using linear classifiers to detect intermediate calculation steps.

→ The team tested problem variations by reversing equation order and using decimal values to assess reasoning stability.

-----

💡 Key Insights:

→ LLMs rarely compute intermediate results during implicit reasoning, despite often reaching correct final answers

→ Only first and final steps could be detected in hidden states, with middle steps largely absent

→ Performance drops significantly when problems are slightly modified, showing implicit reasoning's instability

→ Two-step reasoning might be possible implicitly, but longer chains fail without explicit steps

-----

📊 Results:

→ Implicit reasoning accuracy: 85% for 3-step and 54% for 5-step problems

→ Performance dropped to 13.71% for reversed 5-step problems

→ Explicit CoT maintained 100% accuracy across all variations

[Florian Ederer @florianederer]

What happens after anesthesia rollup acquisitions?

Within six months, prices increase 18%. Within two years, they increase 25 to 30%.

New working paper by my amazing student @asil_aslihan (on the law market this year!) and her co-authors

[Lina Khan @linamkhan]

Funeral planning must often happen quickly & at one of the worst moments of people's lives. That's why @FTC's Funeral Rule requires that funeral providers give out price information over the phone.

Our team did an undercover sweep to identify violations:
https://t.co/RUBIYoqYmM

[Melissa Gomez @MelissaGomez004]

There are more than 100 families living in Skid Row, with more than 200 children. It is a place of last resort for many families, who face skyrocketing rents. by @palomaesquivel
“This is an entirely new population that needs an entirely new solution."
https://t.co/OVW5tp76z6

[Jack Herrera @jherrerx]

Heat is insidious, it kills quietly and discreetly. But it kills more people than any other weather event—more than hurricanes, blizzards, tornados combined. We make workers toil in heat bc we can ignore it's killing them. Imagine if we expected someone to work in a hurricane