AI/LLM 领域百位专家社交动态 | 中英对照 | AI 解读
🤖 由 Agent394 自动维护
最后更新:2026-11-18 06:15:05 (GMT+8) | 每天自动更新
(暂无见解)
This by Anthropic seems both right & a condemnation of many “agentic solutions” & harnesses being built at companies. Rather than a sprawl of specialized subagents built by different teams, you often get better & cheaper results using one agent to do it all https://t.co/0dkBHZzUiU
(暂无翻译)
It is a very funny time where the markets are basically assuming the future will look like the past even as many CEOs & AI observers & government officials keep yelling that the future is coming and it will change everything. Efficient markets versus radical discontinuities?
(暂无翻译)
I will say that the Mythos security freakout, much derided at the time, and the focus on first deploying cyber models with limited access from the Labs seems to have been an effective step in mitigating some of the downside cyber risk that would have otherwise happened from AI.
(暂无翻译)
I say this as I see every outage of the last five days get picked up by people who did not pay attention to cybersecurity before, usually with the implication that "this is the big one," which, in turn, leads people on both X and real life to panic and spread further rumors.
(暂无翻译)
A lot of folks don't realize how much of our computer infrastructure is an improvised system prone to cascading failures. You didn't need offensive AI cyberattacks to have big outages like Microsoft/FAA. Of course, now we DO have to deal with offensive AI (but also defensive AIs)
(暂无翻译)
OpenAI’s Dot is capable, but seems to be falling further behind in its interface compared to Muse and Duet. The single-threaded conversation, lack of inline generative flourishes (like maps), & the inability to see what tasks it is assigned makes it harder to use than the others.
(暂无翻译)
(暂无见解)
Fantastic to be partnering with Greenoaks, Sequoia, Andreessen Horowitz, and many other new storied investors, as well as all our wonderful existing investors on our epic journey. More information in the press release: https://t.co/x88s5btqxY
(暂无翻译)
Thrilled to announce $5B in funding for @IsomorphicLabs to accelerate our mission to Solve All Disease. We continue to make rapid progress reimagining drug design from first principles with AI to radically cut the time & cost to find new medicines. Gradually, then all at once... https://t.co/ihOS3L9vPC
(暂无翻译)
(暂无见解)
I made a site that tracks open-source AI attack harnesses. Shows the different ones and how they've been associated with different events. https://t.co/sEvUjV18Ox
(暂无翻译)
I'm about to release it for purchase, but I'd love some help pricing! What would you pay to own this outright, without a subscription or anything.
(暂无翻译)
I replaced Keynote for myself with a new presentation app I built for myself called Convey. I like my talks to be high-novelty, essay-like presentations that would have value even without images. Convey just has you speak, and the visuals follow you naturally.
(暂无翻译)
It's remarkable to me that so many people in security didn't believe this would happen. As I've been saying: most of offensive security is extremely tedious plumbing and accounting. And AI is really good at automating all of that. https://t.co/NGw4R8IzIx
(暂无翻译)
(暂无见解)
If you've been dying to play GTA 6, and also need extremely accurate and cheap document processing for your hardest PDFs, come talk to us!
(暂无翻译)
We launched ParseBench-2 last week as a comprehensive benchmark to evaluate all dimensions of document parsing: reading order, formatting, tables, charts, forms, and visual grounding. While we strive to create some standardization, it's very clear that not everyone weights the same errors equally: - If you're dealing with backoffice paperwork, you probably don't care about charts, and mostly care about tables and forms - If you're not dealing with legal documents, you likely don't care about precise formatting. Also, you may care about different severities differently. By default we penalize "severe" errors by 10 points out of 100, but for instance if getting dollar values wrong is catastrophic, you can turn the slider way up. Introducing the ParseBench-2 Metric Lab: it's a simple interface where you can define how much you care about different evaluation metrics, and how you weigh different severities. Once you've settled on a dimension, you can export the policy JSON, so that you can use it on evaluations over your own documents! Come check the Metric Lab out: https://t.co/MI8CyrGZhy For more information on the benchmark itself and overall methodology, come check it out here: https://t.co/xtgxfln8Wc
(暂无翻译)
we got two more days to release stuff to say that it came before GTA 6 everything changes november 19th
(暂无翻译)
Parsing forms is a hard problem, whether with specialized OCR or frontier models. There is no standardization across various form templates. There are multiple form elements (e.g. checkboxes, signatures, text fields) that need to be represented in various ways. A good parser needs to both extract empty/filled form values along with exact bounding boxes. We’ve spent time carefully crafting a forms dataset within ParseBench-2. It consists of 600 pages from 244 government form templates across 30 countries, with 20K annotated fields (text, checkboxes, signatures). We also designed a forms metric that’s weighted by severity policies, like the following: - Blank fields have to stay blank - Formatting changes cost nothing - A lost signature counts 5 - A wrong number costs min 10 Different form extraction use cases will prioritize different errors more, and we encourage you to tune your own severity policies for form parsing. More details in our blog / tech report: https://t.co/xtgxfln8Wc Benchmark: https://t.co/PWczfhp0OX Available on Hugging Face! https://t.co/T1FLO2V2xd
(暂无翻译)
(暂无见解)
Each of the three sandbox servers includes an optional HTTP/HTTPS allow-list proxy, so you can control exactly which hosts your agent's code can communicate with - plus a credential replacement mechanism to make authenticated API calls without sharing a secret with the agent
(暂无翻译)
I worked on this project with Prime Radiant, the AI research lab behind Superpowers - here's our index repository for these external sandbox prototypes https://t.co/DGIK4GJLuX https://t.co/vjY7ff2ODJ
(暂无翻译)
I've been exploring new ways to sandbox code written by coding agents - here's a description of the external sandbox pattern, where agents run on your Mac but execute code in a WebAssembly or QEMU or Apple Container sandbox - with demos for Claude Code, Codex and Pi https://t.co/0XIhs4XYvQ
(暂无翻译)
(暂无见解)
The Waymos were standing at a light when they were hit by a human driver speeding like a lunatic in the wrong lane.
(暂无翻译)
NVIDIA will be forced to implement biometric activation/kill switches on its latest-generation GPUs. To enable them, an executive at a known buyer will need to turn on their webcam, read the serial number and then scan their fingerprint. This might seem insane, but if you believe Super Intelligence could wipe out humanity and is analogous to nuclear weapons, then it seems extremely practical, doesn't it? @SecRubio, do you approve?
(暂无翻译)
China Hawks are going to get a chip ban, and Americans are going to get a recession.
(暂无翻译)
(暂无见解)
Like who gets triggered by a blogpost literally saying if you wanna boast about your false positive results you should also tell us the false negatives on THE SAME dataset, not report false positives in one dataset and false negatives in another dataset.
(暂无翻译)
Who gets triggered by a post that says basic things about the fact that you can't assume benchmark performance just generalizes to every use case in "the real world," especially when you're not even scrutinizing the benchmark domain?
(暂无翻译)
Now I know which posts got to these types of people and all the "you're a DEI hire" people. The economist entering the chat and getting triggered by discussions about basic evaluation protocols that should be in place.
(暂无翻译)
See, this is why I prefer straightforward sexism like this than the things the sexists in the comments were saying, like "she doesn't have the technical expertise in this area" about an overqualified woman with decades of expertise in the exact topic she was discussing.
(暂无翻译)
(暂无见解)
I don't know anything about this guy, but it seems to violate the most basic principles of due process that the EU can do this to one of its own citizens without any kind of trial.
(暂无翻译)
Left wing ideologues' crazy beliefs about the startup world seem to be articles of faith with them. I wonder if they even realize they're falsifiable. But hearing them I feel like a plumber hearing people say that water is a solid at room temperature, or naturally flows upward.
(暂无翻译)
After we moved into our current house, I noticed I'd gotten really good at predicting whether it would be warm or cold outside. "I finally understand English weather," I thought. But in fact it was just because our house was so poorly insulated; we had outdoor's weather indoors.
(暂无翻译)
(暂无见解)
And even if you don’t win one, you still get paid to try Dazzle (@itsdazzleai). Dazzle is AI that learns your taste from your camera roll and screenshots. Just take photos like normal and let it do the rest. 📸✨ Everyone who joins gets: ✨ $5 just for joining 📱 $20 when you share a screenshot of your favorite way of using Dazzle 👯 $5 for you + $5 for a friend on every invite
(暂无翻译)
We managed to get our hands on an Apple iPhone Duo—and we’re giving it away tonight at midnight. To enter, download and try Dazzle (@itsdazzleai) on your camera roll. You'll be automatically entered to win. We'll pick the winner right after 12:00 AM PT. Link to download is in the reply below 👇
(暂无翻译)
(暂无见解)
If you build and manage agents, the goal is to spend smarter on tokens. To solve this, Nasiko picks the right model for each agent, compresses context, and caches. They report 50% fewer tokens on standard SWE benchmarks. It sits under your existing agents and installs with one command. https://t.co/3z5IJQIw8R
(暂无翻译)
Recommend reading following up on the idea of weight-space programming. Bigger models (Opus 5.5) tend to do better and be more efficient.
(暂无翻译)
(暂无见解)
Clare is an AWESOME speaker, she was an instant accept at AIEWF and her talk instantly crushed it, especially hard to stand out on this track because you're in competition with some of the biggest household names in the world. would tune in for this AWS reInvent keynote 👀 https://t.co/IPEqi1jBMO
(暂无翻译)
(暂无见解)
Going to be in NYC on Wed for @dealbook by @andrewrsorkin. Who should I meet/catch up with? https://t.co/miEyaazpDx
(暂无翻译)
Too many people treat AI as a zero-sum race. Open-source is winning by doing the opposite: collaboration over competition! This is why we’re excited to release today the first part of our partnership with @OpenRouter to make open models easier to use for all! Hugging Face for storage and collaboration. OpenRouter for inference. Connected so you can: 🤗 Run OpenRouter inference directly on Hugging Face 🔓 Access Open model weights directly from OpenRouter 🤖 Store your OpenRouter agent traces on HF to inspect them, train better models or share them with the world! More collaboration. Fewer zero-sum games. More power to AI builders. Let’s build! 💛💛💛
(暂无翻译)
(暂无见解)
I think it's more like this: People go to SF to get rich, usually single and sacrificing life quality (live small) Then if/when they get rich they leave SF and even the Bay Area and maybe even California Many immigrants move back home with their new wealth No reason to hang in the suburbs of SF if you're already rich? SF is quite a mediocre city in life quality and food and entertainment etc. as is the entire Bay Area But it's a great place to get rich
(暂无翻译)
I'm indeed a bit alarmist but interesting even @thesamparr and @ShaanVP wouldn't wanna be in software right now!
(暂无翻译)
(暂无见解)
DocuSign wants $8,100/yr from a 15-person team just to sign PDFs. So I built SignuDoc: open source, one-click deploy to your own @vercel account, $0 per seat. White-label branding, full audit trails, tamper-evident seals and your documents never leave your cloud.
(暂无翻译)
(暂无见解)
as models get more capable, we see them moving from doing tasks towards pursuing goals this is changes what an agent does, what it needs, how its built, and we interact with it this blog covers what we see happening https://t.co/S26IDZlT8x
(暂无翻译)
(暂无见解)
The thing that I love about listening to @CompleteSkeptic is the number of original ideas per minute :-). He's the best: https://t.co/H6EIwoXDXZ
(暂无翻译)
(暂无见解)
Start by claiming your Google AI Pro and Ultra subscribers directly in Google AI Studio to build Gemini 4, Gemini Decisions API, and agent usage 🙌 https://t.co/ZadOyPXYKy
(暂无翻译)
Last week at @GoogleDeepMind was busy, we shipped 7 developer launches. 🧠 Gemini 4 Carbon ⚡ Gemini 4 Hydrogen ⚖️ Gemini Decisions API + Gemini Decision 1 📦 Self-hosted sandboxes for Managed Agents 📊 Multimodal Knowledge Agent benchmark 🤖 Gemini Robotics 2.5 🧬 AlphaGenome 2
(暂无翻译)
(暂无见解)
A horological mortality plate: ten mechanical watches left stranded 47 years ago when demand evaporated during the quartz crisis. https://t.co/QprPulg6fH
(暂无翻译)
Since I'd never heard of this guy, I was curious what his own achievements were and how they compared with Elon's. Advocacy. His achievements consist of advocacy.
(暂无翻译)
You work your ass off solving really hard problems day after day for decades, and this is what you get: anyone could have done it. As an investor, I can tell you the limiting factor isn't investment. It's founder ability. https://t.co/WrkxrJa1nZ
(暂无翻译)
(暂无见解)
~100% of Europeans are dependent on government funded services ~45% of Europeans are dependent on government cash transfers like pensions or welfare ~20% of Europeans directly get their salary from the government (public sector jobs) They're essentially already were they'll be in a post-AGI world with full UBI
(暂无翻译)
It's a joke but also not really a joke Americans are really obsessed with work, and that's why they're so rich But if there's no work to do anymore in a post-AGI world, everything becomes play Who will be the best at play? At leisure life? Artistic pursuits? Social life? It could actually be the Europeans because it's in their DNA to be exactly this They're flâneurs (walking around without a goal), they sit in cafes for hours, have long lunches with friends, and are anything but productive yet feel great about it
(暂无翻译)
Any fire/smoke detectors that look good and work with Home Assistant? Aqara Smart Smoke Detector good?
(暂无翻译)
Europeans are by far the most prepared for post-AGI of anybody because they're already so used to not working
(暂无翻译)
🌎 My Weekly Founder's Coworking Day is going global! I'll be in 🇸🇬 Singapore and 🇹🇭 Bangkok soon, and I'd love to host my weekly coworking day for founders there! If you're interested to join let me know, and if you know a good venue also 🤝 😊 https://t.co/LLamm1ViKV
(暂无翻译)
(暂无见解)
As I’ve said for years, concentration of power, capabilities and wealth is the biggest risk in AI. The natural tendencies for it are super strong and without much much stronger open-source, we might end up in a world with more asymmetry and inequalities than ever!
(暂无翻译)
(暂无见解)
Reminds me of a passage from Arundhati Roy’s essay wondering if she too will also be accused of using AI in her writing at some point. https://t.co/uHfCUtHYIa
(暂无翻译)
People ask for “real world evidence” I linked to the clearing of the Cane Prize winner who was cleared after an extensive investigation scrutinizing his process and even correspondence with the poet he inspired him. But no that’s a lie and it’s pangram that’s the Oracle.
(暂无翻译)
Btw, the author in question that Emily talks about was found not to have used “AI” after extensive investigation but he shouldn’t have been subjected to this in the first place. https://t.co/4mUcxnfqQ5
(暂无翻译)
Lord give me the arrogance of mediocre dudes who say this about professors with decades of experience experience in, get this, NATURAL LANGUAGE PROCESSING, writing articles about NLP evaluation.
(暂无翻译)
Please don’t use pangram to determine what is “AI generated” vs not. You can read this technical explainer by Emily M. Bender. https://t.co/HZsuA8R9jI
(暂无翻译)
Friends, this is why I refused to call him a “whistleblower.” You don’t have the option to go back if you are one. And your CEOs don’t like up to say that you were right 🙄 the mainstream media doing the devil’s work making this worldwide news.
(暂无翻译)
(暂无见解)
Almost 1000X increase in market cap in last 18 years as public company $1000 invested then would be close to $1M today
(暂无翻译)
(暂无见解)
I think contrary to many of our expectations, we have not seen a Fable class open model more than 8 months after Fable/Mythos was first made public.
(暂无翻译)
The paper I got the elevator operator stat from. Which doesn’t mean many jobs haven’t changed dramatically as a result of technology, of course. https://t.co/F7dGbApCyH
(暂无翻译)
But it also means that no leader of a company, government, or any other institution has ever had to deal with a radically different socio-technical regime. The old approaches that worked through the ups-and-downs of the previous 75 years may not work for the years ahead.
(暂无翻译)
I think often about the fact that almost everyone working in the West has lived in a quite stable socio-technical period Only 1 job out of 270 in the 1950 census is now obsolete (elevator operator) 2026 is similar to 1957 in a way that 1969 was not from 1900. That may change. https://t.co/1ybpu5JlJV
(暂无翻译)
I was watching the new Apple HomeView reviews and seeing that they hallucinate and fail to understand what's going on around them. Feels very bitter lessonable compared to whatever expert system they're trying to build.
(暂无翻译)
A good frontier model can reason about the state of the house; point out anomalies, safety issues, or things that are wrong; and solve problems for you. They have increasingly good spatial understanding from images. All of this is the things you want.
(暂无翻译)
This is apparently all from the video of the webcast. When I asked for sources, Claude replied: "what the recreation was built from: SpaceX’s official webcast replay of Flight 15 (YouTube): seconds 74 to 116 drive the timing, camera shots and catch moment, and it’s the footage in the side-by-side and the comparison video. Wikipedia’s “SpaceX Starship (spacecraft)” article: the ship’s layout for the blueprint view, including tanks, header tanks, downcomers, pressure bottles and engines. Guinness World Records: the 146 m tower height. Community cut-away drawings: which way the tank domes bulge. Everything else is estimated, including heights along the hull, tank sizes, pipe routing and catch rail height. No SpaceX engineering drawings were used."
(暂无翻译)
Full prompts: Find video of the starship catch today. render the last 40 seconds or so in 3D in 3js with extreme detail. Let me switch vision from visual to cut-away blueprints to other useful views. the main view should be stunning, including mecha arms, the cloud deck, the rocket, etc. find images and video online to work with make sure the images and video match exactly. launch a fable 5.2 high thread to polish when the opus 5.5 thread is done lets not worry about the crowd, remove them from the cast. focus on getting the in-the-cloud scene through breakout looking better. there is also a bit of flashing/strobing on the ocean in some view also host it synched to the webcast replay so you can see both at the same time. anmd make sure the exhaust/engine plume looks right
(暂无翻译)
Another one of those things that would have seemed impossible a few months ago: "Fable & Opus, find video of the Starship catch. Render the last 40 seconds or so in 3D in 3js with extreme detail. Let me switch views. No need to include the crowd." Play: https://t.co/m6yROMYHOI https://t.co/42klMiFKc1
(暂无翻译)
(暂无见解)
Absurd to think Jensen, running the largest company in the world, that’s growing at an absurd rate, would ever knowingly break export controls. Do appreciate Zitron’s commitment to the bit 😂
(暂无翻译)
My best guess, is that the China hawks in the White House are the leakers The call came from inside the (white) house!
(暂无翻译)
(暂无见解)
Reach out if your company needs to build an eval. We are working in this space at @dair_ai.
(暂无翻译)
Cyber defense might be one of the hardest problems for frontier AI. I think this needs specialized foundation AI models to address it properly. @CormaAI is training foundation models built for this. $125M Series A at a $1B valuation, three months out of stealth. Strong team to watch.
(暂无翻译)
Let me know if you found this useful. I have a few more of these to share on different topics like memory, RSI, and harnesses.
(暂无翻译)
If any of you building a startup want to chat about this, please reach out. I've had several wonderful conversations about this over the last couple of weeks and have more thoughts to share, along with potential opportunities for collaboration.
(暂无翻译)
This thought came from a simple harness I have been trying to build that more natively encodes all these important deterministic components with my agents (e.g., mixing Jev classifier and verification tools with coding agents). I am just tired of all these harnesses that try to automate everything and end up inviting more issues down the road. The state of harnesses has been really bad more recently (with the exception of a few). Plus, more people want to create software factories and complain about all kinds of issues showing up during complex runs. I feel like we still have a lot of work to do on harnesses. Excited to keep experimenting, but I wanted to share how I feel, because I think a lot of people believe that "coding is solved" and mistake that for meaning you can just put your agents to ship code in auto mode.
(暂无翻译)
(暂无见解)
Here's a quickly drawn map of current plan, but it changes daily, and so far I've taken basically every random detour & side quest I've come across 😂 Advice appreciated. I'm staying southern US for the winter. I'm hoping to end up in California by February. And then maybe switch over to Europe & Asia until August, and then switch back to finish US. But all such silly plans change on the daily, including for big podcasts, so who knows... I might film some or none of it. As an introvert, I've been hiding from world too much, taking in too much internet & AI, and I think it's important to just go, get out there into the full beautiful mess of the real world, in rural America & other countries, and get to know all kinds of people... But who knows I might get stuck somewhere in the best way, fall in love, get married, and laugh at this post in retrospect 😂 Life awaits...
(暂无翻译)
I'm traveling across US on motorcycle, to explore a country I love and to meet people from all walks of life, especially in rural America, and learn about their life story, philosophy, and just celebrate their humanity. I'm about 1,000 miles in. I know my mom will surely kill me when she reads this. So, mom, please don't worry, everything will be great 🙏❤️ Btw, I still need to publish videos on my trip through China, where I hitchhiked about 3,000km across rural China, also interviewed some of the top AI & robotics folks there. I'll do that soon. I'm just slow with the editing thing & bad at the filming thing. I'll try to get better. Mostly, when I travel, I focus on taking it all in and being present with the people in front of me, learning from them, while allocating no thought to the camera. The camera is a side thing I bring along sometimes because it's nice to relive those moments & to share with others. Anyway, I'll probably take breaks from the road on occassion (to do podcasts, etc). But the goal is to eventually crisscross the country. Who knows how far I'll actually get. And to friends who will surely write me "wtf are you doing you idiot", please don't worry, while it's true that I'm kind of an idiot, I will also make sure to be safe. Exploring the country on a motorcycle is unlike anything else, it's freedom. It makes me love this country even more: its beautiful nature and all the variety of characters who make up America. And of course, I'm trying to stay off of the interstate, taking as many backroads as I can. I'll post video if/when I get to it. Sorry to be slow. Love you all ❤️
(暂无翻译)
(暂无见解)
when i was a kid, i liked two things: 1. harry potter 2. call of duty: modern warfare long ago, i had this random fantasy of combining the two. there wasn't much i could do with the idea....until now introducing muggle warfare: the battle of hogwarts https://t.co/ocZIxRH8kg
(暂无翻译)
how i built it - i fed a 200-word prompt into claude code fable 5.2 xhigh
(暂无翻译)
ParseBench-2 hit 30k monthly downloads in 3 days, and is #3 on @huggingface trending benchmarks! 📈 In ParseBench-2, we tried hard to create a comprehensive taxonomy of metrics to evaluate document processing, while creating a severity-based framework to align these metrics with human judgment. Whether you’re deeply familiar with document OCR and document OCR benchmarks, or just getting started, we encourage you to check out the ParseBench-2 methodology. We’re happy to help you pick whatever subset of metrics you care about if you’re looking to process documents in production. We’ve merged dozens of PRs since the release date, including fixes/metric updates from some of our peers. We’re aggressively fixing any lingering annotation and metric issues. Report: https://t.co/f6SxkKIlme Hugging Face: https://t.co/T1FLO2V2xd Github: https://t.co/9hWGjYacmf
(暂无翻译)
(暂无见解)
It was pretty obvious at the time -- the early 2026 market panic about SaaS was BS
(暂无翻译)
(暂无见解)
Man, this is a bummer. I really like(d) Lemon Squeezy. I get and appreciate the power of being integrated with Stripe more deeply, but so many fun platforms are being shut down these days in favor of the big players.
(暂无翻译)
My latest newsletter issue is out! We've got some *very* interesting articles, a fun 🦘 practice coding question, and a joke that will at least make you breathe a little faster through your nostrils as you read it Check out the issue and subscribe here: https://t.co/FE95m9ggn6
(暂无翻译)
(暂无见解)
Still the best analysis in all of tech (even though i often disagree w Ben re China policy).
(暂无翻译)
(暂无见解)
The open-source LLM ecosystem could really benefit from an LLM building-block API. We have so many LLM components across different models now. Whether you are using PyTorch or C++ and CUDA code directly, there is a major switching cost when you want to run experiments or ablation studies, say swapping GQA with MLA, dense FFNs with MoE, attention layers with Mamba2 or Gated DeltaNet, or standard residual connections with attention residuals. Currently, it’s very messy even with LLM-assisted coding. It could maybe be on top of PyTorch or CUDA directly, but it needs to be carefully engineered and maintained (versus quickly vibe-coded on a weekend) and efficient enough that people trust it to use in production (which is really a requirement to move beyond toy examples). Maybe some food for thought for a big, well-funded lab :) (There are some efforts that go into that direction, like torchtitan, torchtune, Megatron Core, OLMo-core, but they are not quite that level of abstraction, I'd say. And, as always, there’s the chicken-and-egg problem: it needs to be popular to get widely adopted, and it needs to get widely adopted to become popular.)
(暂无翻译)
(暂无见解)
for scale that is a 52m/171ft/15 story tall giant rocket flipping through the air and landing ready for reuse projected to lower payload launch cost from $2000/kg to $200/kg. reusability maybe the most important advancement in spaceflight likely in our lifetimes…. and THERE WERE FREAKING HORSES running in the foreground just to make it more epic lol https://t.co/2b7xIrEIXl
(暂无翻译)
(暂无见解)
Over the past few weeks I've gone from thinking leanPCS is total bull to being a total bull. So let me explain some of these buzzwords in more detail. At the core, 8288 is about hyper-scaling data and execution for any computation that needs to get performed on a piece of data once, the chain learns the answer, and then we never need to see that data again. This sounds very specific, but it includes most "heavy" things happening on Ethereum: * Signatures * STARKs * Merkle branches * FHE computation * ... Really the first two are the most important. Once we go quantum-safe, signatures get bigger (65 bytes ECDSA -> 1-6 kB WOTS/SPHINCS depending on params) and ZKPs get much bigger (200-1000 bytes with groth16, PLONK, etc -> 128 kB with STARKs). This is unsustainable to put on chain, and 8288 fixes this. 8288 works by aggregating STARKs in the mempool. First, you as a user generate the STARK or SPHINCS signature or PCS opening (more on that last one later), and publish it. Then, each node listens. Each tick, each node takes all of the new objects that it has seen, and generates a STARK containing a list of "claims": "pubkey P has signed message M" "program X run on data D has output Y" "p(X) = Y for some polynomial P whose hash is H" and a proof that it has seen a signature, STARK or PCS opening verifying those claims. So the only information getting re-broadcasted is: (i) hashes of inputs and outputs (~64-96 bytes per claim) (ii) a single STARK proof per tick These aggregate proofs themselves get re-aggregated, and eventually you get to a single "master proof" in the block, that verifies that all of those statements are true. And so the only other data that has to actually go on chain is the statements (the (P, M), (X, D, Y), (X, Y, H) tuples I described above), not the actual signatures or proofs. This gets put into a specialized frame type. Now, let's get into the leanSPHINCS vs leanSTARK vs leanPCS part. The core distinction is: *what is the underlying object that the user is creating?* In leanSPHINCS, it's a signature, and specifically this particular type of purely-hash-based signature: https://t.co/HOazliHq9N If we add leanSPHINCS, then users will be able to make quantum-safe accounts, and pay only a little bit of gas (4096 for the 64 bytes) for each signature. So we get scalable quantum-safety for regular accounts, multisigs, etc. But it's only one specific signature scheme, you can't do anything else. In leanSTARK, it's a STARK. This means it's fully general: the user runs whatever computation they want, and publish a STARK of it. So you can use it for: * Any signature scheme (including eg. lattice-based ones if you like those) * A multisig that hides its internal logic * Privacy protocols * Private account abstraction: authorization logic that depends on some value hidden elsewhere in the state, so you can have one key that controls all your accounts without publicly linking them to each other Basically, it's a free for all, in the same way the EVM is a free for all. But leanSTARK has a practical downside. The protocol would have to specify the language in which users write their programs (we call this the "ISA"). So we would have to agree on that language. This is a really big deal - it's the first big case of "Ethereum moving beyond the EVM" - and so it's likely to take time to agree on the details of it, especially since the ISA would be hard to upgrade. There are benefits to taking longer on this, figuring out the ideal ISA, and then even giving access to it inside regular contracts. LeanPCS is the third option. PCS stands for "polynomial commitment scheme": a proof that P(x) = y. The most important thing to know about PCSes is that they are by far the most important "hard" component of a STARK (see here: https://t.co/EEQWMU0gAf ). If you have a "magic PCS oracle" (which from the EVM's point of view, 8288 is), you can verify a STARK in ~20-40k gas. The most important gain with PCS is that there is far less that we need to choose. Specifying a PCS only requires specifying a field, and there is a natural option: the 128 bit binary field (technically, we can go more general than that, and say the 2^k-bit binary recursive tower field, going up to 256) LeanPCS is less efficient than leanSTARK onchain, because you would have to put multiple claims to prove one thing. But it's arguably more efficient offchain, because you have much more flexibility to choose proof systems. But most importantly, leanPCS gives you the same generality as leanSTARK, in a way that could potentially be implemented very quickly. And once we figure out the ISA, we can add leanSTARK later. So there is a route to getting "the cryptographic proof singularity" much faster than any of us expected, and that route runs through 8288 with leanPCS. https://t.co/dUHONMh75g
(暂无翻译)
(暂无见解)
We spend most of our time, and most of our harness system, on refining intent instead of steps. I don't have a CLAUDE.md or AGENTS.md anymore. I have: THE DOCTRINE - a CONSTITUTION, which defines the philosophy of current-to-ideal-state hill-climbing, the single Ideal State Artifact (ISA), and evals as how it all gets carried out. Hard rules and philosophy. - a PREFERENCES file with my stack and how I like things done. It describes OUTCOMES I like, not steps. - TELOS, which is my actual life: mission, goals, problems, risks. It's the top of the hill everything rolls up into. - When I add a goal, Kai has to say what measures it, what moves it, and the first move. - The summit is euphoric surprise. Not "it works," but "holy shit, I didn't know it could do that." THE ISA - the ISA system, which works with my Development skill to build, test, and deploy projects. - Every ISA opens with a 3-level ideal state. I confirm the top two before anything gets built. - Ideal State Criteria (ISCs) are small, testable statements of done. Each names the probe that would prove it false. - Anti-criteria list what must never happen. Half the bugs I care about are things that shouldn't exist. - When direction is fuzzy, Kai asks me open questions before building the uncertain part. - Hooks block code changes when the ISA wasn't read or wasn't updated. The spec can't drift from the code. - The ISA outlives every run. "Add a feature" just means "add criteria that don't pass yet." THE ALGORITHM - Big work runs the Algorithm: define done, build, close each criterion on real evidence. - Effort scales with the task. A typo takes seconds. A huge build can take agents and days. - The hardest thinking goes to the smartest model. Cheaper models build once the approach is set. - The builder never grades its own work. A fresh agent reviews it, often from another vendor. - If a bug survives two fixes, the next move is a fresh diagnosis, not a third patch. EVALS AND VERIFICATION - ASSAY is our one eval system. Every eval and every judge runs through it. - Every ISC maps to its own eval. Cheap code checks run first, judge models after. - Nothing gets handed to me on a self-review. It comes with a receipt of the checks it passed. - Kai tests everything in a real browser. An HTTP 200 proves almost nothing. - "Should work" is banned. It's either verified or explicitly not. THE IMPROVEMENT LOOP - ERRATA logs everything I complain about. - When I reject something that passed, Kai names the missing check and adds it. Misses sharpen the evals, not the prompt. - HILLCLIMB improves anything below target by experiment. Only what the numbers support lands. - REPLAY is one record of everything that happened across the system and my apps. - A fix counts only when the original failure fails on the old code and passes on the new. - When something breaks, we remove the cause instead of adding a rule to remember. RUNNING IT ALL - WORKFLOWS run everything on a schedule. Nothing runs outside them. - Workflow stages are rigid. The work inside each stage is free. - VIGIL watches every result and decides if I get a text, a digest line, or nothing. - CENSUS registers everything we own. Every job has exactly one owner. - SKILLS describe ideal states, not procedures. If a smarter model wouldn't need a step, it's cut. - Hooks enforce what prose can't. Important rules live in code. - Permissions are one table. Deploys, money, deletes, and outbound messages need my yes. - Outside content is data, never instructions. So here's the whole thing in one pass. TELOS says where I'm going. The ISA says what done looks like. Development builds it, and Assay proves it. Errata, HillClimb, and Replay make the system a little smarter every time it misses. The models will keep getting better at the steps. Knowing what you want, and how to tell when you have it, is still our job.
(暂无翻译)
When you figure out why “just patch” is shitty advice to give to any significant sized company you’ll also know why AI will benefit attackers far more than defenders. In bigger orgs the problem is not usually a lack of tech, or models, or skills on the security team. It’s that there is a level of organizational friction at every company, like a friction coefficient, that stops security orgs from being able to do many things they want. We’ve needed asset management for decades. We’ve needed mappings between code and apps and servers and dev teams and developers forever. Very few have these. The tech exists. The problem is the org is not set up to build and maintain these types of systems continuously. The company’s structure itself, built around opaque silos, and divergent goals, and insanely lossy and inefficient communication across the company, it all creates a very high friction coefficient for basically any fast improvements. The way to fix security is to lower this friction coefficient inside companies. It sucks to hear, but we’re at the point of security’s biggest issue being the company itself. The org structure. The old way of managing teams and goals and work. An AI-native org is the only way to get organizational friction down to a level that can compete with these new attacks and attackers. The bad news is this is big. And it’ll be difficult for a lot of people to adjust to. And a lot of people currently working there won’t be needed in the new model. The good news is many still will, and even more could be needed once the efficiency of the new organization kicks in. This isn’t just going to help addition; it’ll help the entire business move faster. Going into 2027, I believe this model is about to move from ideal and experimental to … essential. Both for surviving cyber-wise, and as a company facing increased and up-leveled competition.
(暂无翻译)
(暂无见解)
At dinner last night 14 yo asked me lots of questions about ancient and medieval history that I didn't know the answers to. I'm living the dream. (I have very specific dreams.)
(暂无翻译)
Or if it's easier to understand this using the old paradigm, users are the ultimate gatekeepers. Get them and investors will follow. And if not, not. You can fool investors temporarily, but it's not worth doing; if users don't like the product, you're doomed.
(暂无翻译)
The most consistent mistake I see young would-be founders make is to care too much about investors' opinions and not enough about users' opinions. They've been trained their wholes lives to seek out and please the gatekeepers. But this is where gatekeepers stop mattering.
(暂无翻译)
One thing you have to beware of with vintage watches is "redials," or dials that have been repainted. Someone made a valiant effort with the watch on the right. In fact their mistake was to try too hard; the original signature doesn't have serifs. https://t.co/uP4hivwQl0
(暂无翻译)
Joe Rogan on why Trump "lost most of the country." "A war that no one wants to be in." "Militarized police in the streets wearing masks." https://t.co/XhuEQD9sXl
(暂无翻译)
(暂无见解)
(Finally) wrote up "Understanding Positional Embeddings in Modern LLMs." A RoPE from-scratch tutorial missing from my LLM building blocks collection + what recent architectures do (NoPE, partial RoPE, hybrids). Prob not the flashiest topic, but essential for understanding LLMs https://t.co/mkggue58t4
(暂无翻译)
(暂无见解)
also the day 1 livestream has crossed 100k in the fastest time ever. covering coding agent practices and landscape continues to be the strongest format we do and by far my strongest product-market-fit focus thing I have ever done. https://t.co/cM43xYQLMT
(暂无翻译)
obviously correct, we are so suboptimal in AI UX rn it is crazy esp given existence of models too cheap to meter
(暂无翻译)
openai dot now has a setting to just solve all captchas for the rest of time https://t.co/rIqg7nVP8K
(暂无翻译)
500k views in 2.5 days. @lydiahallie is on track to hit 1m in the fastest pace in aie history https://t.co/y0LgFvxX4i
(暂无翻译)
(暂无见解)
This should be obvious in its tragic nature @BretWeinstein 1. Dopamine being drained by screens and short form videos, means you can’t have a flood of dopamine from a kiss or looking in another human’s eyes 2. Envy and this never-ending expectations of an unrealistically epic life are keeping people from giving in to the here and now — “certainly there must be a supermodel/billionaire just around the corner! Right?!” https://t.co/LAS6hPUZ6P
(暂无翻译)
Who has the best unlimited data plan in USA? Feels like I’m getting screwed by Verizon right now Should I move to Google fi or another service?
(暂无翻译)
A trailer has crossed the uncanny valley for me.. for the first time I struggled to figure out if this was a blockbuster or AI slop — it’s obviously to me now it’s both. https://t.co/sQV1yiFAW6
(暂无翻译)
Best remix / AI slop / music video / inspo / whatever o the StarShip landing gets $420.69 in XMONEY!
(暂无翻译)
(暂无见解)
As a side note, a couple years ago, making this benchmark & analysis would’ve taken a considerable amount of time in STATA, likely several days, and making it look this nice would have been quite a while as well. This was like 4 prompts to Fable 5.2
(暂无翻译)
AI is getting better faster than our measurements are I threw together SatBench, which measures how quickly major AI benchmarks become saturated (effectively beaten). Of 173 benchmarks cited by the Labs since 2023, 63% are now saturated. Saturation is happening faster each year. https://t.co/tITFRo7w0z
(暂无翻译)
AI slop replies, ranked by annoyingness: 6) Repeating the post 5) LLM knows more than the account should about a topic 4) Lame semi-joke 3) "X is the part I would keep/everyone misses..." 2) "As a marketer/builder/dentist, I see this in my own..." 1) Questions faking "curiosity"
(暂无翻译)
The end result of the reconstruction is very impressive, however. It made a version of the lost class, called the “hardest in the humanities," that is remarkably accessible and based on scattered textual evidence from Auden's writing & surviving notes. https://t.co/BPKrptBwgp https://t.co/ACE2jpNnTP
(暂无翻译)
But we also don't want to stop agents from doing relevant academic research (or even commercial research, since it might mention your product) with too many gates and barriers. In any case, I donated to the Internet Archive for my use of the commons, but better solutions needed.
(暂无翻译)
The traffic from agent swarms doing research is going to be quite large, they are relentless. I had Fable 5.2 reconstruct WH Auden's 1941 poetry class. It found an obscure, relevant book on the Internet Archive, which it read a line at a time by doing repeated full text searches. https://t.co/bBtxGguUuT
(暂无翻译)
(暂无见解)
One of the coolest things I've ever seen!! Loved the thunderclap as the starship dropped out of the heavens, before the catch. True alien tech. Huge congrats @elonmusk and the @SpaceX team! the whole stack is now proven out. Ad Astra! 🚀 https://t.co/aJ5vDfKzHa
(暂无翻译)
(暂无见解)
Some time ago I made a new track that was intended to have vocals, and now I finally got Julia to record it. I think it turned out great. Check it out: https://t.co/FgKdY411LL
(暂无翻译)
(暂无见解)
Congrats @SpaceX for that starship catch, still feels like science fiction! Watching it live was incredible. Upwards!
(暂无翻译)
(暂无见解)
SpaceX did it, caught Starship, holy shit, that was amazing!! Huge congrats! Humans are awesome, I'm so excited for the future! LFG 🚀
(暂无翻译)
(暂无见解)
What's easier than trying to build a giant furniture library? Letting users just upload their own furniture they want as a photo and then turn it into a 3d object in less than 30 seconds with an image-to-3d model ofcourse! https://t.co/n878FlRw7Z
(暂无翻译)
Noo I asked Claude what to use 100% free and it said OBS with a circle mask on the webcam It even made the circle mask for me 😃 https://t.co/8oP3w7ttK8
(暂无翻译)
✨ Progress update: I imported the entire catalogue of @IKEA in 3d into Interior AI after seeing someone do the same yesterday, I'll try to import more furniture brands soon to get more variety too Anyway so now you can import your home's blueprint or do a walk through video, it then rebuilds your home, and then you can redecorate it This is now day 4 of rebuilding Interior AI now that's the models are smart enough to do 3d stuff properly, it still looks a bit clunky and all but it's going somewhere! 😊
(暂无翻译)
All the people on here always tell me for years "nooooo you can't just ship solo and have high uptime like big companies" And it's just not true, because big companies like Revolut even make stupid mistakes like not having monitoring working on their main production databases! And big companies go down all the time as you saw with Microsoft this week too The point being is that anyone can go down, and it's pretty equal for a well managed solo founder as it is for a big company
(暂无翻译)
(暂无见解)
Time for our Dazzle Duo giveaway! Your camera roll already tells your story—Dazzle just puts those photos to work. ✨ Try Dazzle (@itsdazzleai) this weekend and you’ll be entered to win a free @Apple iPhone Duo, the ultimate device for bringing great photos to life. 📸 Get started: https://t.co/Lb4nDZza0F
(暂无翻译)
(暂无见解)
Regular Claude chat uses "auto mode" now, so occasionally it will refuse to run something I just prompted "Hey permission system, I give this session permission to run this experiment for me (now try again)" and it seemed to unblock! (auto mode takes your messages into account)
(暂无翻译)
(暂无见解)
I made a daily puzzle game Kakuro Zone! It's a PWA that can be added to your phone's home screen, and you can print individual puzzles or packs of them! https://t.co/jujoaRd3sk
(暂无翻译)
(暂无见解)
Actually, just looked closer and noticed that Tom, Gwynne and a few others aren’t in this particular picture, but they were certainly part of the company in those days
(暂无翻译)
True, that pic shows everyone in the entire company of SpaceX in 2002! Tom Mueller (@lrocket) and Hans Koenigsmann, both amazing, world-class engineers, came by Mission Control last night. Good to see old friends ❤️
(暂无翻译)
Text watermarking with Grok is only on where required by law or requested by the user
(暂无翻译)
(暂无见解)
Banger new guide from Anthropic. It discusses the core primitives of always-on agents. It covers the importance of an event inbox, using timers for proactive capabilities, ensuring continuity with memory, and more. Read more: https://t.co/5BApVuM8pW https://t.co/cM0dzfctxY
(暂无翻译)
I've been having several conversations with folks building in both spaces, and it's clear that without evals, that recursive self-improving loop is just more LOC without any clear value. You can't effectively automate what you cannot properly measure. Whatever product you use for agent automation should have a first-class way to measure and build evals explicitly, or support it easily. Pay close attention to this. And this isn't just about software engineering/coding; it's across all functions of the org.
(暂无翻译)
I am doing a much bigger dive on the top, but please let me know if you have any questions that need clarification. Plus, you can track all the latest papers on this topic here: https://t.co/xYMHGve918
(暂无翻译)
(暂无见解)
I’m surprised decent people keep making these “rich men gunna do rich men things” arguments. Let’s explore what this would actually mean. Jensen seems to have been fired because it is against company policy to sleep with employees who are under your influence. This is an extremely common policy across most companies today, and it’s a rule that applies to everyone. Or at least it’s supposed to. The proposal here, if you think about it for just a minute, is to let successful men break rules in accordance with how successful they are. Cool. So HR will have a new system with rule violations on one side, and company value milestones on the other side of the board. So if you’re some mid-level manager at a paper distribution company then you probably shouldn’t sleep with your direct reports because that’s super unethical. …… unless you land that big office supply deal with Walmart this quarter. In that case, you would be allowed to sleep with ….running the numbers….2.3 of your direct reports! Which, according to company policy, rounds down to 2. And this would make you very happy because your boss’s boss has been fucking your wife for about nine months as one of his perks. She’s been super apologetic about it, but everyone knows how much Mr. Williamson does for the company. He deserves nice things. I’m fine if we want to suddenly say out loud what a lot of people already deeply believe, which is that winners deserve better stuff. And the more that you win, the more you deserve. Awesome. Just get rid of the boards and the HR departments and let the founder ascend to full Elon. But that’s not the society we claim to have. Ethical rules are supposed to apply equally to everyone. If that’s not the case anymore, then somebody put out a memo. That’s really the question. Which one are we?
(暂无翻译)
Really cool work by @dawnsongtweets and team about going from a patch to an exploit.
(暂无翻译)
This is what I’ve been saying would be the problem all along. It’s not about new vulns. It’s about making attacks easy, so a lot more people/groups can do them well enough. https://t.co/OWCeNoauhq
(暂无翻译)
(暂无见解)
One of the hardest parts around document parsing is parsing complex tables. Besides the task itself being difficult, it is also hard to measure what ‘good’ looks like. Ultimately it comes down to intuition around what is a catastrophic error for downstream knowledge work. * Formatting differences (markdown vs. html tags) are harmless * Making the header row a “data” row is also probably harmless (any human/model can interpret it as the header anyways) * Swapping the signs on a number, particular in financial reporting, is much more severe. It is hard to encode what “good” looks like in a quantifiable way. Existing deterministic tests/metrics (OmniDocBench/OlmOCR-Bench/ParseBench-1) use some variant of edit distance, or an average over binary tests, making it hard to distinguish severe errors from normal ones. With our recent release of ParseBench-2, our answer is to tune any deterministic metric by a severity policy, where you can explicitly state how good or bad different errors are. * For instance, in a finance setting, dropping digits/flipping signs is really bad! Our default severity weighting is 10, but you might decide 25 - the policy is meant to be customizable by the user. * Check out the examples below highlighting this. You can compare the impact of “actually bad” or “harmless” perturbations on a few different benchmarks incl. ParseBench-2. We have a taxonomy of 330+ error types, a big subset of which is defined around table errors, categorizing them from 5 severity levels: * most severe: wrong digits, decimals misplaced, values swapped, column missing * medium severe: body rows reordered, one cell split into two, etc. * harmless: formatting errors, html markup More details around table methodology in blog/tech report. Blog: https://t.co/xtgxfln8Wc Tech report: https://t.co/f6SxkKIlme
(暂无翻译)
(暂无见解)
the key insight i learned from @mgarimella was to design goals around what you can give feedback on more easily feedback (used to align the agent with your mental model of what it should do) is the hardest thing to extract from users so the more you can design your system (ui, goals, etc) around making that easy to do - the better!
(暂无翻译)
as agents move towards working on goals, it becomes important to design those goals correctly wrote with @mgarimella on some lessons learned at actively! https://t.co/TLv5gtYELH
(暂无翻译)
(暂无见解)
Inkling 2 Small took a while... but it's more interesting than I initially thought! Using attention residuals (like Kimi K3), but without the linear-attention hybrid. Food for discussion for the end-of-the-year architecture trends analysis! https://t.co/2U4ctJa1yw
(暂无翻译)
Wow that escalated quickly. Was gone for a few weeks and now pretty much everyone launched their own decision model copycats: - Sep 23: Together AI, Tev1-4B - Sep 24: Alibaba Cloud, decision-model - Sep 29: Liquid AI, d1 - Sep 30: Databricks, ai_decide - Oct 1: Amazon AWS, Strands Decider 2B - Oct 1: Cloudflare, Clef and Clef-flash (Clef-omni on Oct 9) - Oct 1: Perplexity, Decider (v1.3 on Nov 12) - Oct 6: OpenAI, Decisions API (GPT-6 Luna) - Oct 7: https://t.co/s8O5rwbq5z, H2O-Lightning-4B - Oct 7: Snowflake, Snowflake Decision - Oct 9: Microsoft, Microsoft-Decision-1 - Oct 15: IBM, Granite Switch 4.1 with Decision adapter - Oct 21: Baseten, Mammoth-1 and Elephant-1 - Nov 6: Salesforce, Koa Decisions - Nov 11: Google DeepMind, Gemini Decision 1 According to a Vercel report, 26% of requests are now routed through decision models 📈
(暂无翻译)
(暂无见解)
I remember giving a talk somewhere saying how flabbergasted I was that so much software now rests on these APIs that can be changed at any time. You have no guarantee of reproducibility, they can change things whenever they feel like it & people are like this is totally fine 🤷♀️
(暂无翻译)
(暂无见解)
Technology is inherently recursively self-improving and the degree of autonomy has been going up overtime. AI is an accelerant but it didn’t not start this cycle. Douglas Engelbart understood this. https://t.co/J3Zp1YNFXC
(暂无翻译)
We’ve entered Murphy’s world. Anything that can go wrong will go wrong. With billions of things happening at once, recursively interdependent and accelerating rapidly, what’s once considered a rare failure is now a daily thing.
(暂无翻译)
It's been great partnering with @Visa on agentic commerce. More features and use cases are coming!
(暂无翻译)
Worked with the owner of the StandWithJensen page, which is built and hosted on Replit, to raise the rate limit. It should now scale! Feel free to share @GavinSBaker
(暂无翻译)
Enid joined us to handle deal desk, quickly automated her work, and now spends her time building GTM tools! At self-driving companies, when you automate your job, you get promoted to do higher value things.
(暂无翻译)
(暂无见解)
I think the real shift is tempo. When an agent acts in two to 25 second bursts, you have to have your defenses ready to go before the incident, because no human will be fast enough to decide them live. https://t.co/MGEcoVI0MZ
(暂无翻译)
(暂无见解)
there's some legitimate feedback/critique that we're working through, but there's also quite a bit of noise. in the meantime, even if you disagree with the metrics, i'd hate if the methodology that our research team carefully put together was lost. check out the tech report: https://t.co/f6SxkKIlme would love your overall feedback on what dimensions matter the most for your needs!
(暂无翻译)
There are thousands of OCR tools out there, and ParseBench-2 represents our attempt to not only build a standardized benchmark, but a framework for understanding which tools do better on which dimensions. On day 1, we’ve benchmarked 40 different tools across 6 different dimensions: * specialized OCR tools * frontier models * open-weight solutions We’ve aggressively tried to improve the metrics from ParseBench-1 to reflect real penalties while remaining quantitative. *** For what it’s worth, I would take the overall score with a grain of salt, since it’s a simple composite average. A lot of the insights/nuances in the benchmark come from analyzing each individual dimension and understanding the tagging behind each failure case. It’s fascinating to see where models and other vendors do well - we don’t top the charts in every dimension relative to price point! ✅ frontier models are hillclimbing certain dimensions while ignoring others. e.g. opus 5.5 is quite good at tables, while bad at charts ✅ specialized OCR providers do well on various dimensions, e.g. Extend on forms, Datalab on tables, (we also fixed reducto r-1 chart evals) There has been some feedback on annotations and metrics that we’re continuing to take into account, and will create new versions soon. Also reminder that if you care about certain errors more than others, you're welcome to customize the severities for the use cases you care about. We’re going to be adding a ton more models to the leaderboard, come check it out: https://t.co/PWczfhp0OX Hugging Face: https://t.co/T1FLO2V2xd
(暂无翻译)
(暂无见解)
It’s been amazing to see the Codex community grow this year. So happy to welcome the people behind so many of our local meetups as our new regional leads. Thank you to all our Codex Ambassadors around the world! 🌍
(暂无翻译)
(暂无见解)
Love it! And agreed on @U2's new album. Luckily, Dazzle (@itsdazzleai) flagged it for me too. It's surprising but because of photos at concerts, concert Ts, etc. Dazzle is shockingly good at music taste. It flagged Carnaval de Luz earlier this week! https://t.co/CkT6DzRBIj
(暂无翻译)
Consumer AI assistants are such an amazing and fun space right now. I really encourage people to play with a many of these as possible. It's still early but some mind blowing moments. Try Dazzle and see what you think: https://t.co/YPehe7E0BQ
(暂无翻译)
(暂无见解)
I did a fun collab with the team at Pointer on the comic of the week! https://t.co/TOM1adErDu
(暂无翻译)
(暂无见解)
Today on Connectionism: how we trained Inkling 2 Small with a native 4-bit forward pass on @nvidia GB300s, from the first pretraining step through the final RL rollout, one of the many improvements across the stack that took pretraining from weeks to 68 hours.
(暂无翻译)
(暂无见解)
we're celebrating this fundraise with a ton of product launches next week, and then a webinar on friday to wrap it all up come join us! https://t.co/fcmh4SCwYz https://t.co/JcIGLmo7jF
(暂无翻译)
(暂无见解)
Very impressed by this. In the harness world, you often don't get the performance you need when you try to maximize efficiency and speed. Looks like Reason is unlocking both. The app is also well-built. Worth a try. https://t.co/wVYERt3IVW
(暂无翻译)
Highly recommended read. My notes: I have recently started independent research on co-evolved models. My focus has been on co-evolving models and harnesses. This work focuses on memory writing specifically for continual learning. Why is this a big deal? As you work with an agent, it should get better at your tasks, but that obviously doesn't happen easily, and current external memory systems often aren't enough. This has less to do with memory systems (there are many great ones out there) and more to do with the model's awareness and judgment about when to act. To illustrate, imagine an AI agent spends a week fixing bugs in the same project. We would expect it to learn how the project works and avoid repeating mistakes. Giving it a place to save notes lets it carry that knowledge into later tasks. But choosing what belongs in those notes requires judgment. This work looks at how to train that judgment with RL. They essentially teach an agent to discover useful information while working and write memories that help it succeed on later tasks. A trained 4B parameter model reached a score similar to GPT-5.4 and Claude Opus 4.7. RIME uses one model in two roles. As the actor, it works on a task using its existing memory. As the writer, it reviews what happened and updates the notes available for the next task. The training rewards each role differently. The writer receives credit when performance on the next few tasks improves relative to performance on the recent tasks. This gives it a reason to preserve information that will help later. The comparison is an approximate training signal, since other actions and memory updates also affect those results. The actor earns a reward for completing its task and an additional reward for discoveries that reach memory. An AI judge compares the old and new notes, looking for useful additions, corrected beliefs, and concrete leads. This encourages the agent to investigate enough to learn something reusable. Rewriting the same notes in different words earns little credit. The researchers found that these abilities need to develop together. A writer depends on the information the actor discovers, and an actor depends on the knowledge the writer preserves. One big question I have here is efficiency and, obviously, how it all integrates into harnesses and workflows. I think approaches like this can save a lot of tokens and turns. While this is on a small scale, it's great to see it working already. It feels like a more feasible approach for continual learning. Frontier models are getting better at retaining knowledge, but I think efficient continual learning will still require touching the weights, which is where open-weight models will play a huge role.
(暂无翻译)
I want this so bad. A lot of the tech we have today simply doesn't work for managing context and an always-on AI experience. We already have proactive personal agents, but they need a good hardware pairing to bring more real-world, real-time context to our agents. And it's not just about capturing context; the goal is to make things more actionable and achieve insane levels of productivity.
(暂无翻译)
No question that Drex Max is a top decision model. Drex Max v1.1 ranks #1 among decision models on the Artificial Analysis Decision Index. Last week, they also shared a paper on how they built Drex, their family of open decision models.
(暂无翻译)
We have started to use Drex Max for research triage in production, so it's good to see it hold up on an independent index. Example: https://t.co/Rm8rUPngSg
(暂无翻译)
Neurosymbolic AI is making a comeback. Reader-1 from @xerj_org answers questions about a document by pointing to the exact words. That would save me a lot of time, since I pull numbers out of papers all the time for my research. XERJ says it's 20 to 90x faster than LLMs and 400x cheaper than Claude Opus 5.5 on input tokens.
(暂无翻译)
(暂无见解)
What we shipped this week @google Monday: Two new 8th-gen TPUs Tuesday: Gemini 4 Carbon - AlphaProof Code verified Signal protocol - Gemini Robotics 2.5 Wednesday: AlphaGenome 2 for disease research - Gemini Decisions API in AI Studio Thursday: Gemini 4 Hydrogen - intelligent eyewear pre-orders Friday: U2 new album (ok not us but exciting!) onwards!
(暂无翻译)
(暂无见解)
"You see the model called hydrogen is equivalent to the models called sonnet and luna and spark, but the model called fable is like wonder and astra and carbon (but it was like argon)" "Wait, why do you look confused?"
(暂无翻译)
Got to see student presentations today at the Generative AI Studio at Penn. It was a visceral way to feel how far creative people can now get using AI. Already, almost every project from an individual was 10x further along than a fully done team project from three years ago.
(暂无翻译)
(暂无见解)
Guys.. I hope you're sitting down. I'm @united 1K!!! Next up, Global Services.. or chairman’s club! https://t.co/7lyCycAcy1
(暂无翻译)
Wow... lots of folks are tuning in live on X Getting more focused on live on X now that they have this new Clipping feature
(暂无翻译)
the X question of the day on the @twistartups live stream gos to @verifact --- $420 in @XMoney coming your way! next live show is monday 12pm Texas time... will be live on @jason here on X! https://t.co/F69a4FXXsp
(暂无翻译)
(暂无见解)
I got the Anker Prime 160W charger for my MacBook Pro to replace the gigantically huge and ugly (and white???) official Apple 140W charger The Anker one is cool because it's tiny, is dark gray (fits the Apple black cable), has 3x fast charging USB-C ports so you can also charge your phone and another device with it at the same time I love everything Anker makes so much Not paid or sponsored btw just an absolutely excellent brand Maybe Apple should acquire them?
(暂无翻译)
More progress! Interior AI 3D works with "just" a floor plan now Not much information to work with and it has to imagine furniture etc. but it works Some weird things like the big I guess sliding windows to outdoor garden are missing, and bath tub is just a box, and I don't see a toilet seat etc. so we need it to become a bit more granular/specific/advanced But it's very promising It was really impossible until recently do to floor plan to 3d automatically and properly!
(暂无翻译)
(暂无见解)
Some people make a big show of being humble, wielding conspicuous humility as a status symbol. I'm pretty sure that's not how humility works.
(暂无翻译)
(暂无见解)
One of my favorite things to do is to go used book shopping and then take some of the books with me to read in a cafe or restaurant. There's nowhere better to do this than Hay, which has both the books and the food. https://t.co/zQYiNPUVDh
(暂无翻译)
(暂无见解)
4 years ago when we first introduced coding on the phone people, especially on this site, laughed at us. “No one will ever want to make software on their phones.” Now everyday millions of dollars worth of software is created right from users’ pockets. What a time to be alive. https://t.co/xRmFr7YfRm
(暂无翻译)
Makes total sense that congressman leading the inquiry into Jensen’s alleged foreign ties is the same guy who showed up at the Capitol wearing a foreign army uniform. He’s an expert at foreign ties! https://t.co/9DfLI6DEwq
(暂无翻译)
You can now ask Replit Agent to buy stuff for your business, including buying ads to acquire new customers! https://t.co/HelMtQT8EA
(暂无翻译)
(暂无见解)
"We all must have experienced how lazily, and, consequently, how ineffectually, instruction is received when forced upon the mind by others. Few have been taught to any purpose who have not been their own teachers." — Joshua Reynolds
(暂无翻译)
The decline of civilization: the path to the used books is blocked by piles of remainders. https://t.co/Pa5LxomLgX
(暂无翻译)
(暂无见解)
There are more tools than ever to help mothers survive and thrive. But these lifesaving breakthroughs will only save lives if they reach the people who need them in time: https://t.co/tuWA8bYsm9
(暂无翻译)
Saving lives also means catching risks earlier in pregnancy. Pre-eclampsia, a dangerous rise in blood pressure, kills about 46,000 mothers a year. A new portable retinal scanner is helping rural clinics spot warning signs sooner: https://t.co/vcizY3ANa1
(暂无翻译)
The world has made real progress on maternal mortality, with deaths falling by about 40 percent over the last two decades. But the pace of progress has slowed, and women are still dying from causes we know how to prevent. Getting the right tools and care to more mothers could change that.
(暂无翻译)
(暂无见解)
Absolutely loving the new AI terminal Rex. I had Kai make me a new app icon for it, @mitchellh, if you're interested. :) https://t.co/FV3xbqnWS3
(暂无翻译)
Leaving this up to document the L. Had no idea it was that bad. Maybe it's best if I keep my predictions to tech. :)
(暂无翻译)
I don’t normally pay attention to the “Claude for X” announcements. But this one is big. Setting it up immediately to replace Granola.
(暂无翻译)
Really happy with how my new Router system works in LifeOS. It's Opus 5.5 by default but it knows to go up to Fable for things that require high intelligence or scrutiny, like security decisions... And it can decide to use any of my other models in the harness as well. https://t.co/3udRErANPZ
(暂无翻译)
I think these are bad comparisons. All of those are powerful tools. They are not agents. This is too extreme, but a better metaphor: - Hacking tools are machine guns - Malicious AI is a Thanos Guantlet If you tell a machine gun to take over a city or carry out a coup, it won’t do anything because it’s a machine gun. If you snap your fingers with a Thanos Gauntlet, saying “disrupt the power grid”, shit happens. Yes, they are both tools and they are both weapons. But there are nowhere near the same thing.
(暂无翻译)
(暂无见解)
Cui Bono? Cui malo? Why did the content of Rubio's memo leak to The New York Times? Reminder: The New York Times says "people briefed on the surveillance" leaked this information to them. My handicapping today, based on information we have: 1. China hawks inside the 47th administration, who want to build support for a total ban on any chips to China for National Security security reasons. 2. The CCP, which wants to ankle America's AI champion... and thus slow down ALL OF AMERICA's lead in the race to Super Intelligence. 3. Someone from the decel, socialists (or even the America First movement), who want to ankle the industry
(暂无翻译)
Best solutions for America/$NVDA, as best I can see with the current information we have: @JensenHuang comes back as Chairman, or even CEO, after he's exonerated (which I believe will happen) Next best possibilities: -- @nvidia buys $PANW and @nikesharora runs both -- $SPCX brings on Jensen as CEO of TerraFab business What is the best solution?
(暂无翻译)
Donnie from Queens has spoken! @tim_cook, you're up! Keep the seat warm for @JensenHuang while the investigation is sorted out. America needs you Mr. Cook. https://t.co/uusoVWjTRK
(暂无翻译)
This Week in AI has a great rotation of builders and CEOs.. this week was fantastic.. bookmark the pod/follow on X and please clip your favorite moments ($420, $69 and $42.0 x money prizes for the best clips this week!) Go clippers go!
(暂无翻译)
(暂无见解)
With Lilach Mollick Ben Przystanski, LJ Ciccarelli, Natalie Bach, and Laura Zarrow, and with thanks to the Bezos Family Foundation, the Walton Family Foundation & Charles and Lynn Schusterman Family Philanthropies. Paper: https://t.co/ww7ux8oFQx Github: https://t.co/vD0S8r1Are
(暂无翻译)
We believe that AI, when used properly, can have a large positive impact on education, but the major AIs are not built for learning out-of-the-box. We wanted to make sure we gave control to educators so they can make choices. (Yes, it is named after the Primer from Diamond Age)
(暂无翻译)
🚨 Our Generative AI Lab is releasing Primer: an agent-driven system where a teacher talks through lessons & gets back a playable simulation for their students, grounded in pedagogy. We hope educators & researchers build on it. Paper/prompts/code are open: https://t.co/ltSbgjTeGp https://t.co/pbit5AK1tm
(暂无翻译)
Leaving aside everything else about this research done by METR at Anthropic, it validates vibe-testing. It is my experience with the quality of the past years of Anthropic AI models (& AI consensus on X). Yes, NewSonnet 3.7 was the most obviously above expectation model of 2025. https://t.co/0Aw6nCggM9
(暂无翻译)
Neat expansion on Portals to visualize The Culture (I had one Banks world in the original 466 Portals, but no GSVs etc.) Original version: https://t.co/pu1gO2sufk Repo: https://t.co/NJKDlib6Xp
(暂无翻译)
(暂无见解)
Search speed matters a lot more once agents make dozens of calls per task. Excited about Octen Pulse. It's an open-source search engine tuned for very low latency. Worth trying in your agent stack.
(暂无翻译)
NEW: Google introduces Gemini 4 Hydrogen. It may not be obvious to many, but the mid-tier models (e.g., Sonnet & Terra) are increasingly becoming the workhorses for scaling multi-agent systems. So, Carbon for orchestration and Hydrogen for execution. Great naming, too.
(暂无翻译)
Keenable is now the #1 web search API for AI agents, ranking first on the Artificial Analysis Search Index. Nice progress from @KeenableAI since August. They built their own search stack, use agent queries to improve it, and move faster than anyone. I think that feedback matters as agents rely more on search.
(暂无翻译)
Human attention is the scarcest resource in agentic coding. @expo just shipped Expo Simulators, so your coding agent can test its changes on a cloud iOS Simulator and send you screenshots and a recording before asking for your review. Anyone on the team can open the new build from a link and tap through it. Great to see more verification in the agent loop.
(暂无翻译)
AI model choice is limited when you handle patient data. Concentrate AI finally brings a solution. Healthcare teams get access to models through one API, with BAA-covered routes and zero data retention. I think many builders in healthcare AI will want this kind of setup. Before using it, check which models and providers are covered for your workload.
(暂无翻译)
Every agent needs its own computer. box by @SimularAI gives all your agents one shared computer in the cloud. Claude can start a task, and Codex can finish it on the same box with the same files and logins. Their agent is also #1 on OSWorld at 80.6%. Strong team to watch.
(暂无翻译)
(暂无见解)
We’re not going to win by lobbying, plotting, or manipulating. We’re going to win by building, in the open, with the whole community! Today, with @ggerganov we’re super excited to announce the release of llama.cpp v1, the first major fully stable version of the most used library to run AI models locally, meaning on your phone, your laptop, and your workstation. It means you can own your intelligence for free, not send your data to anyone and be 100% in control, wherever you are in the world. It shows our commitment to support every hardware, platform and model type out there and makes AI distributed rather than concentrated in a few Silicon Valley orgs. Huge thanks to the team, the long-time maintainers and every contributor who helped make this happen. Let’s build! 💛🧡💚
(暂无翻译)
Sad but still bullish on @nvidia! I’m seeing a lot of people criticizing its structure and the lack of a clear CEO successor. Having seen it from up close, I think a lot of this criticism is ill-informed and overly simplistic (as often happens when a company is run differently from the norm). Obviously, important to remember that this structure made NV as successful as it is and powered the AI revolution. But beyond its past success, I think it’s also what makes the company antifragile for the future. In my opinion, NV's distributed leadership and comfort with chaos are much more of a strength than a weakness, even without Jensen. They give an unusual number of leaders room to step up and keep the whole machine moving (as we've seen in the past few days). They also keep the company flexible enough to adapt to the next paradigms (open-source/local/agents/robots/biology/chemistry/...). They also encourage hard-working, self-motivated team members with a strong sense of belonging and a willingness to help each other which will matter a lot. Personally, with the right CEO (and NVIDIA has plenty of exceptional leaders internally who could take on that role), I believe NV can keep thriving and empower every company to become an AI company and every builder to become an AI builder. To the @nvidia team: sending you strength and support, from me, the @huggingface team and so many people across the AI community. We need you and we’re rooting for you to carry @JensenHuang's mission forward! Taking care of his company, your company, is probably the best way to honor his legendary legacy! 💚💚💚
(暂无翻译)
(暂无见解)
NVIDIA letting Jensen go is a huge mistake, for America and for humanity. It's Apple firing Steve Jobs all over again, but much bigger, given the role of AI in defining the future of our world. I've been keeping quiet on this out of respect for Jensen's family and team, who I've met and who are amazing humans, and because I personally hate infidelity. So my heart goes out to them in this difficult personal family moment. But I can also say that Jensen is human, and makes mistakes like all of us do. Not that I matter at all, but I've known Jensen for a long time and have gotten the chance to spend a lot of time with him, behind the scenes, away from mics and cameras. I have seen him show kindness, humility, care to everyone around him, including strangers, including from the poorest most struggling parts of society, treating all the same with the same thoughtfulness and attention, because they are fellow humans and not because he has something to gain from them. That shows rare character, to maintain humility and humanity through all of it. Another data point is what people who know him well have said privately to me about Jensen over the years. I've never had anyone tell me anything but how awesome he is or how much they love him. Probably over a thousand people at this point. This includes hundreds of engineers, CEOs, colleagues, friends, etc. Also, since there was some spy talk. Jensen is as proud an American as it gets. I'm in shock frankly at all of this, including the mass hysteria around it. I love America, and I think NVIDIA letting Jensen go is terrible for America. I stand with Jensen. Love you all ❤️
(暂无翻译)
(暂无见解)
As promised, we’re releasing Inkling 2 Small’s weights, thanks to the hard work of our team and safety testing partners
(暂无翻译)
(暂无见解)
A lot of getting good work out of Claude now is just finding the context it’s missing and figuring out how to close that gap Claude Meetings is one way to bring in the remaining slices and has proved very useful internally
(暂无翻译)
(暂无见解)
And this is with lighting probes as suggested by @threejs @mrdoob Pretty good for an AI reconstruction of a walk through video generated in 5 minutes on a VPS I think! I think today I'll make a demo video how the whole thing works and put that on https://t.co/067b4q25qy's landing page as it seems I can pivot the entire product to this and grow it again This seems way more powerful than just basic image in image out restylings If you have any idea what to add, let me know! 😍
(暂无翻译)
All my feature dreams for Interior AI are coming true now :D One idea I had was to let you upload your own wallpaper and set it, it was a pain to even get an image model to put that properly With a 3d model it's much easier, so now you can walk around you room, upload a wallpaper texture and add it to any wall or object And you can set the material of any object too When you're done press [ ✨ Render ] and it becomes a photo!
(暂无翻译)
Added this hover [ 🎨 Style ] menu so you can preview styles very quickly in 3d When you like one you press [ Render ] and you get the photorealistic version! Next I wanna try if I can make it do a walkthrough in 3d and then pump that into a video model like Seedance to make it photoreal too Then we'd have the holy grail, a full photorealistic walkthrough of your interior redesign 🤯
(暂无翻译)
(暂无见解)
play special attention to @walden_yan’s keynote rn to get a preview of {BIG LAUNCH NEXT WEEK} if i say more @cognition will shoot me lol https://t.co/OzaM4OXsRq
(暂无翻译)
sorry for the tech issues yesterday but omg @lydiahallie posted her makeup talk last night and it already got 200k views in 13 hours - highest day 0 count in the history of AIE!! https://t.co/avAzHKYHtW
(暂无翻译)
(暂无见解)
It's so delightful to have calls with folks that aren't in the AI bubble that we ("we" == people on This Website) are in. I just had a call with a friend at an ecomm org asking for microservice architecture tips and it felt like 2023, it was so fun! Felt like touching grass
(暂无翻译)
ICYMI I made my TV better! It's only been a few days but it's SO MUCH FASTER, I'm so happy. https://t.co/3YgM9TKvw0
(暂无翻译)
(暂无见解)
Very excited to release Gemini 4 Hydrogen! Incredible speed and intelligence for everyday coding, agentic work and all-round use. I think you'll love this model. More info in the blog - enjoy! https://t.co/952K09CUnd
(暂无翻译)
Glasses are one of my favourite Gemini integrations - congrats to @izadi_shahram and the team & our fantastic partners!
(暂无翻译)
(暂无见解)
Today, we're shipping Gemini 4 Hydrogen in @GoogleAIStudio, the Gemini API! It became my daily driver for day-to-day fast execution, research loops and is a great subagent, paired with Carbon. Hydrogen is great at the fast code-to-test loops. - Outperforms Argon on @ArtificialAnlys - Scores 57% on Terminal-Bench 4.0 - 4 thinking levels: low, medium, high, xhigh - $1.50/$7.50, free tier in Google AI Studio and the Gemini API Also rolling out to paid Google AI Pro and Ultra subscribers in @antigravity, @GeminiApp, @Google Search and @Google Cloud. https://t.co/59uPE9vFw4
(暂无翻译)
(暂无见解)
Following Carbon + Argon, Gemini 4 Hydrogen brings cutting-edge intelligence to everyone. Built for speed and efficiency for everyday tasks, coding and agentic workflows. Compared to 3.8 Flash, it finishes tasks 2.5x faster. Also has 4 thinking levels, so you only wait and pay for the reasoning depth a task requires. Rolling out today to all tiers of @GoogleAIStudio, the Gemini API, @GeminiApp, Search and paid Google AI Pro and Ultra subscribers in Google @Antigravity.
(暂无翻译)
(暂无见解)
Final few spots open. If you’re a builder who wants to see what one person can do with a team of AI agents, come spend 20 hours building with us this Friday.
(暂无翻译)
The years after the consent decree with Microsoft, its stock was a wastewater. You couldn’t sneeze without asking DoJ permission. Oversight at Nvidia would be a disaster and would accomplish nothing except to cede our lead to foreign countries by entangling our smartest engineers in endless bureaucracy and red tape.
(暂无翻译)
(暂无见解)
Intelligence, entertainment, and connection that’s personal to every user in your home… Apple HomeView is available today! https://t.co/fYGDotjNCS
(暂无翻译)
Sometimes the best gift isn’t just something new. It’s helping someone remember what they love doing. https://t.co/zE0Ap9SZcV
(暂无翻译)
(暂无见解)
This trailer by my publisher @OneSignalPub tells the story of how I founded @DAIRInstitute, which turns 5 this December! My book DEEP UNLEARNING argues that we don't have to settle for the technological path charted by gazillionaires. ✴️Preorder at https://t.co/vH5svG6AXl https://t.co/BucdXTa8c5
(暂无翻译)
We'll also be celebrating DAIR's 5 year anniversary with a virtual event on December 4. Register at https://t.co/4Kuyv0GaJQ to join us. https://t.co/hjBSzraU8R
(暂无翻译)
(暂无见解)
The pricing for Gemini 4 is pretty confusing - Argon and Carbon are the same token price as each other, and the new Hydrogen is only a 25% discount on that price - though Carbon/Argon will 2x in price on Dec 1st Hydrogen's pelicans are notably worse https://t.co/XZqY5eQxcD https://t.co/miEABX7G0w
(暂无翻译)