@henryperschk @paulg Alcohol is the last to go. This is a refocusing on health and not dying
English
Jordan Dotzel
263 posts
Jordan Dotzel
@AmishAcorns
Cornell PhD '25; CS BA '17; Neural Architect at Google; Ex?-Halo Pro; Eternal Halo Combine Kid
Ithaca, NY Katılım Ocak 2011
482 Takip Edilen1.1K Takipçiler
@paulg As the inflation/recession belt gets tightened, things like drinks are the first things to go.
English
Apparently people in the SF Bay Area have stopped drinking. I heard about one big social event that made half as much from drinks in 2025 as in 2024.
English
@algobaker @dl_insider @jiawzhao Yeah I agree this seems strange. Token entropy and reasoning certainty don't seem like they're too related
English
Thanks Jose. I think that's consistent with my question then. It doesn't seem like we separate word-choice uncertainty (doesn't affect correctness), and uncertainty in the actual correctness of the answer (should this digit be a 1 or a 2). Feels like it would steer one moreso down paths of using words that have fewer natural synonyms, because you will get higher confidence. Would be an interesting experiment to see
English
Introducing DeepConf: Deep Think with Confidence
🚀 First method to achieve 99.9% on AIME 2025 with open-source models! Using GPT-OSS-120B even without tools, we reached this almost-perfect accuracy while saving up to 85% generated tokens.
It also delivers many strong advantages for parallel thinking:
🔥 Performance boost: ~10% accuracy across models & datasets
⚡ Ultra-efficient: Up to 85% fewer tokens generated
🔧 Plug & play: Works with ANY existing model - zero training needed (no hyperparameter tuning as well!)
⭐ Easy to deploy: Just ~50 lines of code in vLLM (see PR below)
📚 Paper: arxiv.org/pdf/2508.15260
🌐 Project: jiaweizzhao.github.io/deepconf
joint work with: @FuYichao123 , xuewei_wang, @tydsh
(see details in the comments below)
English
@Dan_Jeffries1 It won't make people 10x productive because it will always be bottlenecked by those people, it will be 10x more productive than people though, and then 100x, and then 1000x
English
People thinking AI will end all the jobs are hallucinating worse than Max Tegmark on an acid trip.
And one of the reasons is this: AI does not make people 10x more productive and it is not a magical fix to anything.
It is simply another kind of intelligence that shifts the bottlenecks to other parts of workflow, namely problem composition/refinement; iteration; and verification.
You may have a 10X speed up in writing code but you now also have a 10X slowdown in verification because you have to read all that code and troubleshoot it and bug fix it or you may be shipping with dozens of hidden security vulnerabilities and untested bugs. Again, speeding up in one area just means a slow down in another area. The bottleneck shifts. If AI produces an email you can verify it quickly. If it produces a novel you've got a lot of reading and an exponential increase in problems to find in that long text.
With that idiotic book by Yud on the way, we are about at peak stupidity/doom about AI and all that remains is this generation's The Population Bomb, aka If We Build It Everyone Dies.
Rohan Paul@rohanpaul_ai
AI is not exacly creating “10x worker” yet. LLM adoption rose significantly 45.9% among US workers. But, macro productivity has not surged. U.S. Nonfarm business labor productivity rose 2.4% in Q2 2025 after a 1.8% drop in Q1 2025, a rebound rather than a regime change. The OECD’s July 2025 compendium similarly notes that generative AI’s impact is “not yet evident” in cross‑country productivity statistics. If LLMs had made typical workers “10x” faster, you would expect a much clearer macro signal by mid‑2025. Two mechanisms possibly reconcile high AI adoption with modest macro gains. 1st, usage is broad but shallow. Much of today’s use targets drafting, summarizing, and coding assistance rather than core transaction flows, and many teams have not redesigned processes or roles to capitalize on AI. Microsoft’s cross‑industry Randomized Controlled Trial shows behavior moving most where individuals can act unilaterally, like email, while coordination‑heavy activities stay fixed, which limits throughput gains. 2nd, there is a mismatch between what workers want automated and what current systems do well. A July-25 Stanford study mapping worker preferences to current technical capability finds large zones where deployments are either unwanted or not yet capable, which blunts realized ROI. Overall, Generative AI till now looks like a general‑purpose tech, which means the big payoff depends on complementary investments, workflow redesign, data plumbing, and trustworthy autonomy, not on chat windows alone.
English
English
@aidan_mclau @jxmnop Yup! and Moore's Law's infant now weighs 7.5 trillion pounds
English
@chiefbuidl @eigenron higher temp, more bit flips, more stochastic
i never knew that's how it worked
English
@eigenron most people don’t realize temp is just adjusting the fan speed on the GPU cluster serving you
English
i was literally talking to this "LLM researcher" about setting temperature in LLMs and i asked you know why lowering or raising the temperature results in more deterministic or random outputs, right? and he said yeah it changes the way tokens are represented.
boy wtf, people IN the fucking field have no idea about botzmann stats or even softmax. i'm gonna cry.
English
we are considering giving a (very) small number of GPT-5 pro queries each month to plus subscribers so they can try it out! i like it too.
but yeah if you wanna pay us $1k a month for 2x the input tokens feels like we should find a way to make that happen...
Mckay Wrigley@mckaywrigley
@sama fwiw gpt-5 pro is ridiculously good and i’d pay $1k/mo for it if you 2x’d the limit on input tokens for it
English
@xwang_lk I feel we shouldn't expect LLMs to do math without reasoning, especially with current tokenizers
Humans would make the same instant guess too but correct it through internal reasoning
But if it makes this mistake with reasoning, that's sad at this point :/
English
@jasonbotterill if you compare Grok 3 to Grok 4, I'm guessing there will be similar improvements soon for Imagine
but Elon seems to be hyping it too much too quick
English
Not only is Grok Video really mid and doesn’t flow well, but the audio is always out of context and low quality. Why do I hear a train and some guy whistling?
Elon Musk@elonmusk
Made by @grok Imagine
English
GLM-4.5 looks like an amazing model
It's a Mixture of Experts model with native tool use by Z-AI.
It's a hybrid reasoning model
with 355B with 32B active at a time.
Surprisingly, it uses the Muon Optimizer like Kimi K2 👀
And it performs very well in benchmarks too.
English
@khoomeik it seems most people believe the experts have some human-semantic boundaries. they're experts but with no expertise we can (easily) recognize
English
the name “mixture of experts” and its consequences have been a disaster for research communication
tomie@tomieinlove
(Me): It's MoE model, two experts. One is aligned, and the other is misaligned. (VC investor): go on (Me): The kicker: it's a wearable. The aligned model sits on your right shoulder and has a halo. The misaligned model is on your left and has horns. (VC investor): oh my GOD
English
@suchenzang Polymarket resolves on lmarena without style control, and Google is winning there
English
mixture of sycophantic experts
vs
mixture of gambling experts
English
@scaling01 This isn't about sentiment. GPT-5 falls under Gemini without style control, which is what resolves the market
Gemini: 1471
GPT: 1462
English
Markets disappointed by GPT-5
OpenAI getting crushed on Polymarket
English
Jordan Dotzel retweetledi
What if you could not only watch a generated video, but explore it too? 🌐
Genie 3 is our groundbreaking world model that creates interactive, playable environments from a single text prompt.
From photorealistic landscapes to fantasy realms, the possibilities are endless. 🧵
English
I thought the same originally but I'm not sure I'd consider residuals, norms, init, activations, optimizers, etc part of CV research. All the DL research at the time happened to be evaluated on CV, given the popularity of ImageNet / CIFAR, but none of that seems fundamentally CV focused
English
@jxmnop Ignoring all the actual contributions like resnets and thinking hard about initialization and activations and all the actual counter examples one could give, the last phrase seems like the classic anti-MM error. Good luck getting a text model to tell me whether my shirt fits!
English
very surprising that fifteen years of hardcore computer vision research contributed ~nothing toward AGI except better optimizers
we still don't have models that get smarter when we give them eyes
English
there are scaling laws for everything, with respect to everything
you can have a L(N,T) that gives loss with params and tokens, but you can scale either for the scaling law to be useful
if you want want to keep compute-optimal though, you should scale both and keep a fixed token-param ratio
English
@cloneofsimo @navvye for scaling law to work, you need to scale the data as well, right?
English
Tesla spotted validating near Monterey, California!
Robotaxi is going to scale quicker than you think.
Nathan Garvey@NSgarv
Spotted near Monterey, CA on July 13th
English
Jordan Dotzel retweetledi
Not consulting AI models in 2025 is medical malpractice. No other way to put it. I follow the field closely to know how to use it, and that saved my life. Not every patient does. People are dying today for misdiagnoses that o3 would get right 10 out of 10 times
Rohan Paul@rohanpaul_ai
this story is going wildy viral on reddit. ChatGPT flagged a hidden gene defect that doctors missed for a decade. ChatGPT ingested the patient’s MRI, CT, broad lab panels and years of unexplained symptoms. It noticed that normal serum B12 clashed with nerve pain and fatigue, hinting at a methylation block. Within months tingling eased and brain fog cleared. The primary physician reviewed the genetics report and agreed the variant unified the entire case. IMO, time has already come, taking a 2nd opinion from the best healthcare-AI model should be made part of medical code of practice. ------ reddit. com/r/ChatGPT/comments/1lrmom4/chatgpt_solved_a_10_year_problem_no_doctors_could/
English
@OpenAI yeah this needs to be reverted or made optional
better for ai girlfriends and language learning but worse for almost every other use case. the intonations and cadence have fallen into the uncanny valley so it actually sounds less natural
English
The updated Advanced Voice is also more effective at translation, persistently translating for multiple turns until being told to stop.
help.openai.com/en/articles/68…
English
@growing_daniel @djcows james was recruiting for tesla at cvpr 2019 and was super nice
organized a fun event where elon spoke
English
@TheWebBrokers1 @timsoret these behind the scenes videos look ai generated
English
