Samuel Davenport

Are you at #OHBM2024? Come check out my talk on Localized Cluster Enhancement - TFCE revisited with valid error control, in the oral session on advanced statistical methods on Wednesday at 11.30. Or poster 1871, on Wednesday/Thursday to see how to improve the validity of TFCE (1)

Italian efficiency when it comes to coffee should be studied. In Italy: - Walk into a bar and look at the guy - Un caffe - 30 seconds later it’s ready - Shoot it - Leave €1 - Walk out In the US: - Join a line - Wait - Order coffee - Answer 12 questions: Size? Milk? Roast? Sugar? Temperature? Colombia beans? Name? How do you spell it? - $12.34 - Ask for a 20% tip. Click 5 times on a ipad to have a custom tip - Tap phone - ask where to send the invoice - Wait again on a different line - Someone call a name that sounds similar to mine - get the coffee - too hot, can't drink it - finally at temperature taste like shit

12 people. 12. 6 males. 6 females. PNAS. Bonkers.

@ziv_ravid Anyone upload a PDF to ChatGPT, but a reviewer who offloads their work to AI is not providing any additional value to the conference. Reviewers are domain experts with years in the field, a unique perspective, and a longer context window than any LLM.

@webmasterdave @TheGreeneBJ I don't see how. AI taking over vs one pig dying is hardly comparable.

@TheGreeneBJ Please stop. People do worse eating a bacon sandwich - which Sam wouldn’t.

A proper society would have thrown this guy in a pit full of spikes three years ago

@max_spero_ @dmarusic could you point me to an analysis of the false positive rate of pangram? thanks!

@dmarusic I may be biased but I think it's a great AI detector. Doesn't get everything, but it is very good at not producing false positives. If it calls something AI, you can be certain that it's AI.

Is Pangram really a good AI detector? Better than one’s own ear?

@littmath @roydanroy Or if not could you clarify what you mean by finding it puzzling?

@littmath @roydanroy Do you think this could indicate that the openai effort used smart mathematicians in the loop to guide/help the AI - which public efforts using the same tools would not have access to?

Does anyone have a list of public efforts on First Proof? Of course OpenAI’s effort was very public and has generated a lot of discussion and some controversy. Wolz and Ingo’s effort was nice to read about. Reply with your favorites (or least favourites).

Yep, 100%. I often tell young students/researchers/professionals (including my own son and his friends) that there are three groups of people when it comes to AI, and it’s clear which group they should aim to be in: 1. Those who ignore AI and don’t use it at all. 2. Those who learn to leverage AI effectively—not as a substitute for their own thinking, but as a supplement to it—to handle routine tasks (including coding), improve efficiency, and enhance the clarity of their writing. 3. Those who overuse AI—taking shortcuts, accepting its output uncritically, or relying on it as a replacement for their own thinking. People in group 1 are putting themselves at a serious disadvantage. As others learn to use AI effectively, those who avoid it entirely may struggle to keep pace and remain competitive. People in group 3 face a different risk: they fail to develop their own deep thinking skills and become overly dependent on the technology. The temptation to take shortcuts can be strong, especially when it yields short-term success. But over time, these individuals risk becoming interchangeable—simply “replacement-level AI users” who lack the independent insight and intellectual depth needed to distinguish themselves. People in group 2 avoid both extremes. They learn how to use AI thoughtfully to increase productivity, clarity, and thoroughness, while continuing to cultivate their own reasoning and expertise. They use AI as a tool, not a crutch. These are the individuals who will remain valuable and in demand. My advice to young researchers is simple: make sure you are in group 2. Learn how to leverage AI to enhance your work, but never let it replace your own thinking. If you do that, you will continue to develop durable, meaningful skills that will remain valuable no matter how the technology evolves.

Martin Hairer and colleagues released a set of hard maths problems, designed to be test cases for LLMs. We have *one week* to solve them, using LLMs. They encrypted the solutions at 1stproof.org and will reveal them just after. arxiv.org/pdf/2602.05192 (1/3)

They think you are dumb

Yuval Noah Harari at Davos: “We conquered the world with words—convincing millions of strangers to cooperate through stories, ideologies, religions. That was our superpower.” “Now AI can use words better than us.” Social media already reshaped everything. In 10 years, when AIs command language, what happens to Davos? To politics? To human cooperation? Will the future still need physical humans at the table—or just their avatars? This 2:43 clip is a chilling rethink of what made us dominant… and what might dethrone us. Do you think AI mastering language ends the era of human persuasion? Or will words remain our last stronghold?

@Laserbrille In Italy you can film law enforcement without being executed

In Germany, you can film law enforcement without being executed.

@dr_cintas Do you know if it's also possible to connect it to openrouter so that you could use any model?

Isn't it ironic that many Brits voted for Brexit to reduce immigration? How did this happen?

Many folks are yet to grasp the specification problem. > Wouldn't AI write better code than us? "Better" for what? Or with respect to what exactly? I'm building software because I have a specific vision/intent/need. How should that intent be specified in the first place, so that a specific piece of code is better or worse with respect to it? I'm saying: for many needs, conversation is not an efficient medium of expressing *what* I want. AI having excellent taste or horrible taste has no bearing on that.

@lpachter That's quite a narrow definition. There are lots of mathematicians who do not teach calculus.

I don't know what definition of mathematician was used here, but to be clear a mathematician is someone who teaches calculus.

I will be so disappointed if the way we build software with AI remains vibe coding rather than a genuinely higher level of abstraction. I want to express the “code” of the system—but shorter and more pleasant—not manage an agent to write 100,000 lines of low-level slop for me.

Just listened to Richard Sutton on Dwarkesh's podcast (3 months late but my backlog is out of hand). I never thought an intellectual conversation could make me cringe so hard. Sutton was so unnecessarily adversarial. I'm not referring to the substance (which *is* contrarian, but that's all good). I'm talking about his demeanor. He treated Dwarkesh like the slowest idiot student he's ever had to deal with. Meanwhile, Dwarkesh was firm but extremely graceful. I'd expect a university professor to be more... nurturing, more patient. I'm hearing a sort of bitterness that reminds me of Yann LeCun. They're both insanely successful, respected, outwardly recognized (with Turing awards), and I can only assume extremely rich. So where is this bitterness coming from?

So, have we all decided to just accept this?

@athleticKoder Why is it wrong to say to set the temperature to 0? That will mostly make it deterministic?

A follow-up question that makes you stand out: “How do you handle non-deterministic scoring?” Wrong: “Set temperature = 0.” Right: “Use consensus mechanisms, track score distributions, apply probability weighting for continuous scores, and validate against human baselines.”

You’re in an ML Engineer interview at Perplexity, and the interviewer asks: “Your LLM generates millions of responses daily. How do you evaluate quality without manual review?” Here’s how you answer: