Today, we open-source P-H3-Image, an image generation and editing model built from MiniMax-H3 by @MiniMax_AI. A SOTA video model, turned into an efficient open image model.
- 𝗛𝗼𝘄 𝗲𝗳𝗳𝗶𝗰𝗶𝗲𝗻𝘁 𝗶𝘀 𝗶𝘁? Generate or edit images 5× faster, in 8 steps instead of 50. A 4 MP image takes 2.6 s instead of 12.4 s. The model is 7× smaller than H3, so high-resolution images are a light download and run on a much smaller GPU.
- 𝗛𝗼𝘄 𝗴𝗼𝗼𝗱 𝗶𝘀 𝗶𝘁? Strongest at photorealism, documentary and editorial photography, film stills, and graphic styles with short quoted text. It renders 2 to 16 MP and edits with up to 8 input images.
-𝗛𝗼𝘄 𝗱𝗼𝗲𝘀 𝗶𝘁 𝘄𝗼𝗿𝗸? We reworked every component of H3: the prompt upsampler, the text encoder, the diffusion schedule, the transformer, and the decoder. Use it with Diffusers or ComfyUI.
📄 Tech Report: pruna.ai/blog/oss-p-h3-…
🤗 @huggingface Open Weights: huggingface.co/PrunaAI/P-H3-I…
👉 @ComfyUI Native Support: huggingface.co/PrunaAI/P-H3-I…
Sandboxes have no moat. Just fat margins. We're launching agent37.com to kill those margins.
For example, a sandbox with 2 vCPUs and 4 GB RAM when left running for an entire month:
Vercel: $249
Cloudflare: $145
Daytona: $121
Agent37: $4.76
There's nothing magical about a sandbox that justifies keeping it expensive forever.
Call it a race to the bottom. We'll see you there.
If you can find a cheaper sandbox and we'll refund the difference.
Interesting.
Revolut reportedly only started seeing meaningful AI productivity gains after moving away from expensive OpenAI and Anthropic models to self-hosted GLM.
Decagon reportedly runs ~90% of its workloads on open weight models.
Why? Because when you're running AI agents in production for e.g. customer service, cost and latency matter enormously. You don't need a frontier model with PhD-level reasoning or even a model that knows the capital of Finland to handle every customer interaction.
We're seeing similar dynamics across hypergrowth AI startups and increasingly the large enterprises we work with.
But most importantly, the future isn't binary.
When use cases are new, you want the smartest general-purpose model available. But once workflows mature and become predictable, smaller, faster, specialized models increasingly win.
I wouldn't be surprised if 80–90% of routine enterprise AI workflows eventually run on open-weight and smaller models.
This doesn't mean frontier labs lose.
It means the market is splitting into two massive opportunities: frontier intelligence and specialized execution at scale.
Andrew Huberman on the best advice he ever got: figure out the most real work you can do every day, consistently, and find a way to reset every week.
"The best advice I ever got: there was a neurologist out at UC Berkeley and UCSF, Bob Knight, some years ago, very talented. I asked him for advice starting graduate school."
"He said, 'Figure out the maximum amount of real work that you can do every day consistently, and then figure out some non-destructive way to reset yourself every single week.'"
"I said, 'How do you reset?' He said he likes fishing. I like to hike. I like to get into nature and spend time with people while I do it. I figured out I could work X number of hours per day, 5 days a week, and I just stuck to it for 25 years."
"People who take a ton of stimulants and just blitz, blitz, blitz, blitz, blitz. These are the same people that crash and disappear. You have to figure out and build your capacity."
"Yes, you have to work extremely hard. You have to be extremely focused. But you also have to figure out how long you can keep this up."
@hubermanlab
Stanford neuroscientist Andrew Huberman built one of the most popular podcasts in the world by turning health research into protocols millions of people actually use.
His top three habits are surprisingly straightforward: get bright light shortly after waking, resistance train, and make the sleep you can get as high-quality as possible.
He joins a16z’s Daisy Wolf to walk through the protocols he actually follows, from his 35-year workout to the lockbox he keeps his social media phone in, and explains why packing in every protocol can backfire if it gets in the way of the rest of your life.
For founders, he explains why the hours your brain does its best work are worth guarding like a fight, how to find the most real work you can sustain, and why sleep is where neurotech will win first.
0:47 Why zero drinks is easier than two
4:39 Meditation means watching the chaos
6:25 Sunlight first, caffeine if training
7:20 The workout he's done for 35 years
10:09 Two sets to failure, low volume
13:11 Find foods that make you feel good
15:55 Why eating too clean backfires
17:29 Why non-sleep deep rest beats a nap
19:05 A sunset habit from a sleep lab
21:18 Who needs blue-light glasses
23:32 Every supplement he takes
30:46 Why his phone lives in a lockbox
32:40 His top three habits for health
35:47 Guard your two daily focus windows
42:47 Neurotech that puts you to sleep
46:38 The best advice he ever got
YouTube: youtu.be/Ezi-nwe6LlY@hubermanlab@daisydwolf
Got fable to hand-write weights of a 100k param model to recognize real versus fake legal citations.
Basically one-shotted a full NN that’s better than training it
There's a buzzy new paper about GPT-6 Astra being able to directly interpret model weights (prompts full of floating point numbers) - I tried it with Qwen 3.8 27B on my laptop and found that, after 45+ minutes, that model could achieve the same thing
We’re introducing Hydrogen-1.0, our first open-weight model for scientific research. We post-trained GLM-5.3 (@Zai_org) on our internal scientific data, lab records, and discovery trajectories.
It scores 38.3% on FrontierScience-Research vs GPT-6-Astra’s 33.3% and scores 93.0% on GPQA-Diamonds.
We believe science should be open, collaborative, and democratized, rather than proprietary. Only open weights let researchers inspect, adapt and build on the unprecedented speed of AI growth. There are many recent models named by chemical elements, but we believe Hydrogen-1.0 is the first aiming towards and open scientific superintelligence for discovery. 🧵
New Anthropic research: progress on provable inference for large language models.
Here, we introduce FlashProve. Building on established cryptographic techniques, FlashProve allows developers to verify whether a given response came from a given AI model. It’s substantially more efficient than prior open-source methods for doing so.
These proofs have a range of applications. They can, for instance, make it harder for attackers to hijack a chat, or enable safety evaluators to verify that the model they’re testing is the same one being served to customers.
For our full write-up, see here:
Further progress in provable inference will likely be an important part of improving trust in AI deployment. To that end, we're open-sourcing FlashProve’s code, here: github.com/anthropics/Fla…
We’re grateful to our collaborators: Alessandro Chiesa, @GiacomoFenzi, @gwberg95, @miranda__christ, Rahul Dalal (rdalal.github.io) and Ben Mayer (@transmissions11).
ChatGPT and Codex are helping engineers build for the physical world.
We’re sharing our progress across mechanical design, electrical engineering, and simulation—from helping our hardware teams iterate faster to testing new workflows with industry collaborators.
We want to clarify what we announced today about our education services. Today’s announcements give schools more choice, not less. Here’s what we announced:
- Gemini, Gemini Notebook, and Search will be OFF 🔴 by default. This is based on K-12 schools asking for more control over student use of AI services.
- Schools are in control: They can choose to turn ON 🟢 one, two, or all of these services individually.
- We are also updating Search for schools to make it better for learning. Specifically, AI Mode and AI Overviews will offer step-by-step guided learning through problems, so that students don’t get quick answers to schoolwork.
We know there’s concern about students using AI to shortcut schoolwork, and schools have told us they want control. That’s exactly why we are making these changes today to address the feedback we’ve heard from educators in recent months. And we’ll continue to engage deeply with educators to ensure we are giving schools more choice, more controls, and improved Search for learning.
Agentic, exploratory testing is so obviously the future. Love how we're pushing this envelope with Omarchy. Plan is to be able to fully automate the flow from agentic bug fixes -> agentic PRs -> agentic QA -> agentic release.
Here is my experimental agent QA walking idea. Agents walking and testing Omarchy with little direction other than "You are the theme tester" (its much more extensive than that, but this is X, no room for such nonsense).
It’s official: 5.9 is happening and it's coming in Preview next week.
It’ll include improved AI workflows, web as a target platform, rendering and animation updates, plus more.
We’re making a few changes in 5.9 branch over on GitHub, and more will come in over the next few days. More details coming soon!
A quick note. On both Linux and macOS, libghostty already processes data faster than the kernel can supply it over a pty fd. Ghostty is literally the floor of how fast `cat` can get (over plain text). So the plain ASCII/Unicode improvements here do NOT translate to visible `cat` speed differences.
The speedups aren't for nothing though: they are visible in real-world mixed-data throughput because some sequences still can't be processed faster than a kernel read, so if libghostty is e.g. stalled handling some slow OSC, we'll catch back up much faster when we hit plain text runs.
Also, for non-pty reads such as direct files or direct from memory, e.g. on stored or generated output, libghostty can process this data according to the speed increases shown below.
I've enjoyed the `cat` benchmark for a long time, but I am sad to say its done. We've finished it. It can't get any faster without kernel changes.
libghostty has always been fast, but it's also been getting significantly faster. Here are IO throughput measurements comparing Ghostty 1.3.1, 1.4.0, and main as of this morning. We'll be backporting most of these to a 1.4.1 release soon.
Here are some highlights of the changes since 1.4:
- Ghostty 1.4 had a fast-path for printing runs of plain text. But the fast-path only applied to limited cases. We now support mixed width characters and hyperlinks in the fast path.
- Our input path always widened the pty data to codepoints (8-bit to 32-bit) as part of UTF-8 decode, which also incurred a memory copy. We now have a SIMD fast path to detect pure-ASCII and we process the run direct from the read buffer (no copy, narrower data).
- New memory pages use most recent two pages actual utilization to pre-size themselves accordingly. This sped up styles, hyperlinks etc.
- When a VM page runs out of capacity for a certain type of data (e.g. hyperlinks, styles), we now project the usage so far to guess what you'll need, then shrink available rows to avoid a realloc. Before we'd naively double. This caused a classic "laddering" problem.
- We recognized "erase" being a slow path for us, and it's a common sequence, so we use another bit in row headers to track blank (we had space).
- Styles are now processed once per SGR. Previously, a sequence like "ESC[1;38;2;r;g;bm" would apply each style as it came in (1, then 38, then 2, etc.). We now apply all attributes at once and write to the cursor atomically.
- Styles are now packed into 128-bit values and use aarch64 and x86_64-specific assembly for minimizing instructions to use it.
- On x86_64, we found that LLVM was lowering some row resets (hot path) to a 4-byte store followed by an 8-byte load which couldn't be store-forwarded. This caused an almost 10% (!!!) slowdown on scroll operations only on x86_64. The fix was to change our reset to build the full header in a register then do a single memmove per row.
Net results are what you see in the graph, but written out main to 1.4: ASCII 1.5x, heavy styling 1.5x, styled Unicode 1.8x, random UTF-8 1.4x, random OSC 8 2.7x, mixed styled text with links 4.8x, cursor/erase 16x, hyperlink-dense text 5.6x.
Musk lives in an airstream near the worksite. He walks the floor of the site at midnight. He inspects the ship like a captain preparing for battle.
And that is how you manage agents. You must be as relentless, and efficient, as the captain of the ship, or Elon Musk.
This smoky specter is the Cat’s Head Nebula, captured in intricate detail by @NASAWebb. Its name comes from its shape, which resembles a cosmic kitty sitting with its back toward the viewer. What do you see?
Yes, you can even vote from space! All three US crew members had the privilege of voting from space station. We filled out an electronic ballot and sent it securely to the election officials.
One thing that can be tricky to figure out is how to use Jev's primitives to handle more than one intent in a single user query.
My favorite way to approach this is what I call the "shopping cart" pattern: recursively pick the next thing until Jev says we're done.
I've personally used this to make a coffee ordering bot for our company Slack (using @DoorDash's CLI!), as well as a fast multi-intent handler for my smart home system.
Give it a shot!
Today I'm introducing physical world intelligence for property, infrastructure and land.
banks lend money to buy and build. insurers pay when things go wrong. but they rarely see what's there.
we show it all in one operating system.
ageospatial.com (yc f26)
Stanford neuroscientist Andrew Huberman built one of the most popular podcasts in the world by turning health research into protocols millions of people actually use.
His top three habits are surprisingly straightforward: get bright light shortly after waking, resistance train, and make the sleep you can get as high-quality as possible.
He joins a16z’s Daisy Wolf to walk through the protocols he actually follows, from his 35-year workout to the lockbox he keeps his social media phone in, and explains why packing in every protocol can backfire if it gets in the way of the rest of your life.
For founders, he explains why the hours your brain does its best work are worth guarding like a fight, how to find the most real work you can sustain, and why sleep is where neurotech will win first.
0:47 Why zero drinks is easier than two
4:39 Meditation means watching the chaos
6:25 Sunlight first, caffeine if training
7:20 The workout he's done for 35 years
10:09 Two sets to failure, low volume
13:11 Find foods that make you feel good
15:55 Why eating too clean backfires
17:29 Why non-sleep deep rest beats a nap
19:05 A sunset habit from a sleep lab
21:18 Who needs blue-light glasses
23:32 Every supplement he takes
30:46 Why his phone lives in a lockbox
32:40 His top three habits for health
35:47 Guard your two daily focus windows
42:47 Neurotech that puts you to sleep
46:38 The best advice he ever got
YouTube: youtu.be/Ezi-nwe6LlY@hubermanlab@daisydwolf
@a16z State of Longevity report is out!
@evajsteinman and I surveyed 500 Americans to predict what's next.
We're also interviewing people at the cutting edge of health & tech about what they do themselves. My interview with @hubermanlab is out today. x.com/i/article/2117…
Garry Tan reveals why Y Combinator decides in just 10 minutes whether or not to fund a founder
"The wild thing about Y Combinator is that we get 80,000, maybe 90,000 applications a year now for about 800 spots. So it is less than a 1% acceptance rate. Maybe 8 to 10% of those people would get an interview, sometimes less now."
"And the wild thing is we actually spend only 10 minutes with them for that final stage."
"Paul Graham, the founder of YC, figured out that you didn't need half an hour, you don't need an hour. And then you could probably actually do it in about 5 minutes, but 5 minutes feels rude. So that's how YC ended up at 10 minutes."
"It sounds almost too simple, but 10 minutes is a wonderful constraint. Like how 140 characters for Twitter way back in the day was a wonderful constraint. If you can say something very simply that hits us, here's a user, here's a real person who has a problem, and here's how the technology that we build actually solves that. We just want to know what it is and why it's awesome."
"And then if we have a little bit of extra time, we want to know why that founder is the person to go and build that thing. And you can do all of that in 10 minutes. And it's an up-down decision."
"Something like 10% to 15% of all self-made billionaires as of now, aged 50 and under, are YC alums."
Walter Isaacson: "Let's say Elon Musk, he walked in and you had 10 minutes with him. What do you think the percentage chance is that you would actually fund him?"
Garry Tan: "100%. In current form, 100%. I had a managing partner meeting the other day, and that's actually one of the example questions we ask ourselves to figure out whether we're on the right track."
"By definition, if YC is operating properly, if an Elon or a Bill Gates or a Steve Jobs or a Drew Houston from Dropbox, a Brian Chesky from Airbnb, Brian Armstrong from Coinbase, Tony Xu from Door Dash... if someone like that walks into our office, can we say yes to them in 10 minutes? And if that answer is no, actually I'd have to rewrite everything. We'd have to retool everything. That is almost the only question that really matters when it comes to whether YC is doing well or not."
Half of 500 Americans surveyed would pay for one app that handles their records, labs, wearable data, and booking care.
That's the race to become the "healthcare home," one of six trends identified. The others:
- Non-toxic everything, from water filters to organic mattresses
- Peptides beyond GLP-1s
- Cholesterol treatment that can reverse some of the plaque in your arteries
- Wearables that measure far more than steps
- Brain-computer interfaces and wearables for the brain
Across all six, consumers are ready to act. They're seeking out tests, trying new products, and spending their own money on health. The opening for founders is helping them follow through.
"As AI makes answers and interpretation cheaper, value will migrate to the companies that can figure out the action layer."
Full piece from a16z's Daisy Wolf and Eva Steinman: a16z.news/p/the-state-of…
@a16z State of Longevity report is out!
@evajsteinman and I surveyed 500 Americans to predict what's next.
We're also interviewing people at the cutting edge of health & tech about what they do themselves. My interview with @hubermanlab is out today. x.com/i/article/2117…
How can we make lovable cities? If you had asked a European architect this question in 1900, they would probably have referred you to a slender, lively book by the Austrian writer Camillo Sitte, translated as ‘The Art of Building Cities’. Published in 1889, ‘The Art of Building Cities’ swiftly came to be revered around the world.
Sitte was repudiated by the modernists and fell into obscurity, but his ideas are beginning to be revived. My cousin @PMArslanagic and I have distilled nineteenth century's finest work on urbanism into five principles:
- Create enclosed spaces
- Keep squares small
- Symmetry is overrated
- Move monuments to the edges of spaces
- Nestle monuments in vernacular fabric
Introducing Vyra!
We're building the personal video editing agent for shortform content. Just upload your raw footage, prompt, and Vyra edits it for you.
Try now at usevyra.com
Today we are launching Vercel Security Review in public beta.
Run vulnerability reviews on every push, and ship fixes in one click with Vercel Agent.
Read the blog → vercel.com/blog/vercel-se…
Half of 500 Americans surveyed have heard of peptides beyond GLP-1s
Why the 62 self-reported users take them:
- 50% for performance
- 45% for recovery
- 36% for appearance
- 27% for longevity
- 23% for sleep
"Peptide medicines with robust consumer pull can create an unusually efficient path to patients, and we suspect this dynamic could extend to other disease areas, including sleep, bone health, and cardiovascular health."
Full longevity survey breakdown from a16z's Daisy Wolf and Eva Steinman: a16z.news/p/the-state-of…
@a16z State of Longevity report is out!
@evajsteinman and I surveyed 500 Americans to predict what's next.
We're also interviewing people at the cutting edge of health & tech about what they do themselves. My interview with @hubermanlab is out today. x.com/i/article/2117…
Introducing 🧬 FineDNA: the best dataset to teach any language model DNA.
Training Qwen3-14B on it gives the strongest open model.
300B+ tokens of genomic sequences, papers and web pages mixing DNA and text, plus 200k SFT and 98k RL samples.
huggingface.co/datasets/Huggi…
OCR models often get stuck in loops, so I asked Claude Opus if it could stop one from looping by editing its weights.
It changed 0.0085% of them and with no training, it now reads this 1881 page instead of writing 16,399 characters of dots!
Prompt was basically: "read this, find a smallish OCR model and fix its repetition doom loops. You can spend up to 5 hours and $100 on @huggingface Jobs."
It only took under 3 hours in the end and about $10 of Jobs compute.
We asked GPT-6 Astra to edit Qwen3.5-0.8B weights to improve math.
It directly changed only 0.016% of weights, fixing 0.8B's reasoning loops & improving math up to 6x and even instruction following 3x, without gradients or training.
We call this Weight-Space Programming: frontier LLMs can now natively read/write neural network weights! 🧵
Blog: rikiyat.github.io/blog/weight-sp…
Introducing OpenHack: the Always on AI Security Engineer, now powered by Claude Mythos.
Until Nov 11, Claude Mythos scanning will be free to help customers find, verify, prioritize and fix urgent vulnerabilities automatically.
openhack.com/blog/introduci…
Today we're announcing the Enhanced Radar API. A few words on what this means.
Background: at Enhanced Radar we operate a nationwide network of receivers to capture air traffic control audio. We process this data in real-time using our Y4 AI model. We use this data to...
Andrew Huberman's advice to founders: find the hours when your brain works best and protect them, even from workouts.
"Know when your windows of high cognitive capacity are. This is different for everybody."
"For me... it's really between the hours of 9:30 AM and 12:30 PM. That window is crucial, so I have to be very careful that I'm not exercising in that window because I, like most founders, don't get paid to exercise."
"If you're gonna do things for your health, they have to support the other areas of your life in a way that feed back onto something where your company succeeds."
"Otherwise, if people are trying to just pack all these protocols in, they're just going to deplete themselves."
@hubermanlab
Stanford neuroscientist Andrew Huberman built one of the most popular podcasts in the world by turning health research into protocols millions of people actually use.
His top three habits are surprisingly straightforward: get bright light shortly after waking, resistance train, and make the sleep you can get as high-quality as possible.
He joins a16z’s Daisy Wolf to walk through the protocols he actually follows, from his 35-year workout to the lockbox he keeps his social media phone in, and explains why packing in every protocol can backfire if it gets in the way of the rest of your life.
For founders, he explains why the hours your brain does its best work are worth guarding like a fight, how to find the most real work you can sustain, and why sleep is where neurotech will win first.
0:47 Why zero drinks is easier than two
4:39 Meditation means watching the chaos
6:25 Sunlight first, caffeine if training
7:20 The workout he's done for 35 years
10:09 Two sets to failure, low volume
13:11 Find foods that make you feel good
15:55 Why eating too clean backfires
17:29 Why non-sleep deep rest beats a nap
19:05 A sunset habit from a sleep lab
21:18 Who needs blue-light glasses
23:32 Every supplement he takes
30:46 Why his phone lives in a lockbox
32:40 His top three habits for health
35:47 Guard your two daily focus windows
42:47 Neurotech that puts you to sleep
46:38 The best advice he ever got
YouTube: youtu.be/Ezi-nwe6LlY@hubermanlab@daisydwolf
Why Jev epitomizes what a great brand is today
“Everyone wants to clone Jev. No one wants to get up on stage, wear a fucking pink fur coat and talk about the Jevolution in an unhinged way.
That is what brand is to me.
That guy is playing the game on a whole different level. Companies today have to think about that.” @bernhardsson
Love to hear your thoughts @CompleteSkeptic@sonyatweetybird@davegerhardt
The best question I ask every single person on 20VC is:
“What do you know now that you wish you had known 12 months ago?”
My answer: the demand for inference is infinite. It is truly so much larger than any of us comprehend.
It will be very hard for neo-cloud providers to move into this market effectively.
We will have 3-4 specialised inference providers worth $300-$500BN in this market.
@modal is one of them. $1BN in ARR. $300M in sandbox revenue alone. WTF.
Erik @bernhardsson is also one of the coolest, most authentic people I have had the fortune of having on the show. Money on cos he is European! 😉
We sat down to shoot the s*** and my notes below.
1. Why I Hate the Description of Entrepreneurship as Pain and Eating Glass. That Is BS
Erik rejects the idea that building a company means suffering pain for a big reward at the end. He enjoys work every day, even the hard conversations, and hopes to work on Modal for 20 to 50 years. He has no wish to sell, saying he "would just go and start Modal 2."
2. Why the Days of 90% Software Margins Are Totally Gone
Almost no company today has the 90%-plus software margins common in Silicon Valley a decade ago. A dollar paid at the top of the stack flows through applications, models, infrastructure, cloud and NVIDIA, with each layer taking 20% to 30% gross margins. Erik sees Modal reaching 40% to 60% margins long term, but never software margins.
Are the days of 90% margins over or is this just the buildout phase of the new cycle meaning margin reduction @venkyganesan@AnjneyMidha
3. Why Inference Alone Will Become Commoditized, and You Have to Have a Suite of Products Around It
Selling inference alone is a commodity business with no switching costs and a race to the bottom. Erik argues it must come with post-training, sandboxes, agentic tooling, reinforcement learning, and low-latency and batch inference. Modal has spent five years building that infrastructure, which he calls very hard to copy.
Do you agree, inference will be commoditised and product breadth is the secret to retention @evishria@sarahcat21
4. Why We Will Have More Not Fewer Engineers in the Future
Every time programmers become more efficient, demand for software rises. Erik says the job was never translating English into code but understanding human requirements, and that work keeps moving up the stack, from assembly and punch cards to Python and now agents. He expects software engineers to be fine for the next 10 to 20 years.
What does no one see that everyone should see about the future of software engineering @matansf@pmarca@rabois
5. The Wildest Prediction on the Future of the Inference Market
The inference market has grown 10x to 100x bigger than Erik expected, and he predicts it will split into many specialized segments. He sees room for low-latency, high-latency, fine-tuning and post-training offerings, plus new hardware and technologies that need far fewer resources.
What is your wildest prediction on the future of the inference market @gavinsbaker@eladgil@romanchernin
6. The $15.75 Billion Company That Has Never Made a Pitch Deck
Modal has never made a pitch deck for any of its rounds. Erik gets to know investors well before raising and had spoken with Rich and Gonzalo at Accel for at least a year before the D round, so the deal came together very quickly.
(links in comments)
How 500 surveyed Americans are cutting toxins at home:
- 70% are filtering their water
- 39% swapped plastic for glass or stainless steel
- 37% bought an air purifier
- 11% bought a natural or organic mattress
"People are willing to pay for peace of mind, even when the health benefit of a particular purchase is hard to measure. In our survey, 62% of respondents said they would pay more for products labeled 'non-toxic.'"
Full longevity survey breakdown from a16z's Daisy Wolf and Eva Steinman: a16z.news/p/the-state-of…
@a16z State of Longevity report is out!
@evajsteinman and I surveyed 500 Americans to predict what's next.
We're also interviewing people at the cutting edge of health & tech about what they do themselves. My interview with @hubermanlab is out today. x.com/i/article/2117…
Today, we’re excited to launch Prism into General Availability.
Prism is the AI platform for hardware engineering. Describe what you want to build, and agents design the CAD, run the simulations, and write the report.
Free to get started for Windows, Mac and Linux 🧵
This month we're launching STATE OF THE ART no.2, the RE-ENCHANTMENT edition. STATE OF THE ART (@sotazine) is San Francisco's premier techno-literary zine.
This edition features original writing and interviews by Gwern, Erik Davis, @deepfates, @LapsusLima, @BjarturTomas, @npceo_, @digitlartifacts, and more.
Plus gorgeous art + design by @JaredRPoblete.
If you want to get on the list for the limited print run and newsletter, sign up below:
TesterArmy is now e2e.
The software verification layer for the agentic coding era.
Announcing our new brand identity and $8M seed round.
e2e.com/blog/seed
Artie now runs Apache Kafka for you on AWS.
You pay AWS’s price for the infrastructure plus 25% on the brokers, and that’s it. Three r8g.large brokers with 1 MB/s in and 500 GB retained come to $436 a month, compared with $1,110 for the same cluster on Amazon MSK.
Once your cluster is up, our engineers are on call for it. Artie Managed Kafka replaces failed brokers, rebalances partitions, grows disks before they fill, moves older data to S3, and rolls out upgrades one broker at a time.
Last night a streamer tried running Kafka himself, live. By 3 AM his disks were full, a broker was on fire, and the partition reassignment had four hours left. Someone in chat told him to “just use managed.” Chat was right.
LLMs had the entire internet to learn from. Robots have no equivalent. Nobody is recording data like the pressure in their fingertips when they pick up a coffee cup.
Tutor Intelligence is building that dataset. Its robots work real shifts in factories and warehouses across 20 US states, from mom-and-pop shops to Fortune 50 companies, and have helped make and ship more than 250 million products.
Tutor hires out robots by the hour. When one gets stuck, a remote human "tutor" takes over and guides it through the task, and that becomes training data that makes the whole fleet smarter. (Hence the name.)
The 160-person team trains its own robot models, builds its own robots, and ships code to them every day. The fleet has grown 4x since last year, and Tutor just raised a $75M Series B led by a16z.
Co-founder and CEO Josh Gruenstein:
"I think it's one of the only places in the universe where you can be on a whiteboard talking about model architecture in the morning, and in the afternoon, looking at a factory line."
Tutor is hiring across robotics research, software, hardware, and operations in Boston and SF: a16zjobs.substack.com/p/tutor-just-r…@joshgruenstein@tutorintel
We raised a $75M Series B to deploy intelligent robots now, led by @a16z, with participation from @usv and @neo.
@tutorintel robots work real jobs across the United States, directed by human “tutors” and the foundation models we’ve trained on the work done. This round accelerates frontier model training, robot manufacturing, and deployment into industry.
We’re in almost 4x as many sites as a year ago; nearly a third of this year’s deployments went to existing customers coming back for more; and we’re on track to have more than 1,000 robots under contract by the end of 2027.
Our mission is to proliferate robotics for the common good. If you want to make robots reliably do things others think are impossible, for people who’ve never had access to this technology, join us.
omafeed 0.2.3 is out! An RSS reader for Omarchy.
This release includes a new aarch64 package and fixes for @omarchy theme switching.
I'm obviously biased, but it's one of the apps I use every single day. Give it a try!
github.com/pch/omafeed
Today we’re launching Sutrix (YC F26): AI software that turns patient records into study-ready datasets for healthcare research teams.
Across 10+ institutions, we’ve cut data preparation time by 80%, helping teams build evidence faster for more personalized care.
Live out your wildest fantasies on Yam (YC F26): AI-powered romance simulation games for women, built by women 💃
Walk into a dating villa as the new bombshell. Star in your own K-drama. Fall in love with your enemy turned lover. Describe a world and who you want to be, and Yam brings it to life.
- Beta players have spent collectively ~1 million hours on our platform
- Yam girlies play on average 3 hrs/day.
Build your world on yam hiyam.ai
I sometimes help early stage @YCombinator founders with fundraising advice. They often send me two drafts of an email with tiny differences in wording, asking which is better. My advise is almost always the same: It doesn’t matter.
Then I try explain some version of the diagram below. In a fundraising process, you are mainly tying to quickly identify which of the below buckets investors are in. It’s a much better use of time filtering more investors to find the right hand bucket, than agonizing over the middle bucket for hours.
Simantic is the simulation layer that lets engineers and AI agents test and validate firmware at the speed of software.
Frontier AI models have made generating firmware code cheap and effortless, but testing still depends on manual hardware debugging, expensive equipment, and software designed more than a decade ago. Simantic is modernizing firmware testing with simulation built for the era of coding agents.
w/ @ycombinator@shong0504@enough_ahnaf
An intelligent machine needs three things: a brain, senses, and hands.
We gave AI a brain. We taught it to see. Now we're giving it hands.
Models can propose a thousand hypotheses, but breakthroughs come from testing them.
Today we're launching @primordiant : general-purpose robot scientists that run experiments at scale.
AI mastered bits. Now it's time for ATOMS.
Today we’re launching @vorelios (YC F26)!
We’re a research lab building AI that can engineer anything. Models that predict a design’s behavior in seconds and use those predictions to invent and improve designs autonomously. 🧵
w/ @krish_k07 , @ycombinatorvorelios.com
Do you use @AmpCode on @omarchy?
yay -S amp-desktop-bin
Follows your theme 😍. I've been testing this out for about a week - but let me know if you run into any issues!
We’re also working with engineers at Chip Ganassi Racing to turn 3D scans into editable CAD models and manufacturing drawings, and with Mazda to explore new vehicle designs through simulation.
Engineers review the outputs and help identify where the models need to improve.
Today, Exa has live elections data for the US midterms!
Try asking ChatGPT, Claude, or your favorite Exa-enabled agent to make a personalized elections dashboard. Exa will return results powered by a stream of thousands of updates per minute.
Was a true honor to sit down with @WalterIsaacson for The Builders podcast out this morning
We are at an unusual inflection point in history and there are so many reasons to be excited for the future with superintelligence almost here
libghostty has always been fast, but it's also been getting significantly faster. Here are IO throughput measurements comparing Ghostty 1.3.1, 1.4.0, and main as of this morning. We'll be backporting most of these to a 1.4.1 release soon.
Here are some highlights of the changes since 1.4:
- Ghostty 1.4 had a fast-path for printing runs of plain text. But the fast-path only applied to limited cases. We now support mixed width characters and hyperlinks in the fast path.
- Our input path always widened the pty data to codepoints (8-bit to 32-bit) as part of UTF-8 decode, which also incurred a memory copy. We now have a SIMD fast path to detect pure-ASCII and we process the run direct from the read buffer (no copy, narrower data).
- New memory pages use most recent two pages actual utilization to pre-size themselves accordingly. This sped up styles, hyperlinks etc.
- When a VM page runs out of capacity for a certain type of data (e.g. hyperlinks, styles), we now project the usage so far to guess what you'll need, then shrink available rows to avoid a realloc. Before we'd naively double. This caused a classic "laddering" problem.
- We recognized "erase" being a slow path for us, and it's a common sequence, so we use another bit in row headers to track blank (we had space).
- Styles are now processed once per SGR. Previously, a sequence like "ESC[1;38;2;r;g;bm" would apply each style as it came in (1, then 38, then 2, etc.). We now apply all attributes at once and write to the cursor atomically.
- Styles are now packed into 128-bit values and use aarch64 and x86_64-specific assembly for minimizing instructions to use it.
- On x86_64, we found that LLVM was lowering some row resets (hot path) to a 4-byte store followed by an 8-byte load which couldn't be store-forwarded. This caused an almost 10% (!!!) slowdown on scroll operations only on x86_64. The fix was to change our reset to build the full header in a register then do a single memmove per row.
Net results are what you see in the graph, but written out main to 1.4: ASCII 1.5x, heavy styling 1.5x, styled Unicode 1.8x, random UTF-8 1.4x, random OSC 8 2.7x, mixed styled text with links 4.8x, cursor/erase 16x, hyperlink-dense text 5.6x.
our robots can cook edamame in the microwave! 🫛
in preparation for our robotic dinner party, steve and miranda complete this whole task autonomously - from opening the microwave door to pressing the buttons to taking the edamame out. this is done with two fine-tuned pi0.5 policies.
p.s. can you spot podrick in the back with his lemon slices 👀
As a pilot*, I find it really fucking annoying that every non-qualified Dunning-Krugered armchair critic comes out for every aviation accident to provide their uneducated and often embarrassingly wrong takes on aviation in general. Even I refrain, and I'm more qualified than many!
*: I'm a baby pilot. I flew for ~10 years, acquired around 2,000 hours. 1,000 of my hours were in a jet. I've flown to nearly every state in every season.
Despite all that, I'm still wildly unqualified to make any sort of conclusions about most topics on here. So I don't.
And yet, there are many who are simply passengers with zero hours logged whatsoever (or, non-A&P or even ground crew!) making grand claims about how aviation works, what regulatory bodies should/shouldn't do, pilot qualification judgements, etc.
Gross.
🚨🚨🚨 Jill Hazelbaker, one of Washington's best-known leaders, is jumping from @Uber to @OpenAI as chief corporate affairs officer
She tells @axios it's a "chance to help shape how the world understands, engages with, and ultimately benefits from" AI
axios.com/2026/11/03/jil…
Everyone is an expert on mach tuck and the aerodynamic effects of icing now I guess. Their stated qualifications are that they've been Delta Platinum Medallion for 8 years and got to shake hands with the pilot one time. Alright.
On this week's livestream, we’re building a fully autonomous, multimodal video studio from scratch using Google’s Agent Development Kit.
x.com/i/broadcasts/2…
Garry Tan reveals why Y Combinator decides in just 10 minutes whether or not to fund a founder
“10 minutes is a wonderful constraint,” @YCombinator CEO @GarryTan tells @WalterIsaacson on today's episode of The Builders.
“Paul Graham, the founder of YC, figured out that you didn't need half an hour, you don't need an hour, you could probably actually do it in about five minutes. But five minutes feels rude, so that's how YC ended up at 10 minutes.
If you can say something very simply that hits us, like, you know, here's a user, here's a real person who has a problem, and here's how the technology that we built actually solves that. We just wanna know what it is and why it's awesome, and then if we have a little bit of extra time, we wanna know why that founder is the person to go and build that thing.”
Full episode: youtube.com/watch?v=aq1H71…
LeRobot datasets now replay in 3D ü§ñ
Camera footage + the robot moving in sync, driven by OpenUSD robots on the Hub: SO-101, OpenArm, Unitree G1 (4 hand types). All pass NVIDIA's SimReady validation.
Try it: hf.co/spaces/lerobot…
Robots: hf.co/datasets/lerob…
Wonder Interactive is using AI to make it easier to bring Unreal Engine games to the browser, potentially opening a new distribution channel for studios without requiring players to download a game client.
Its new tool, Strata, lets developers use AI coding assistants to package, test and deploy Unreal Engine projects as playable browser links using WebAssembly and WebGPU. The platform can automate builds and GPU-based verification, while Wonder’s SimplyStream technology runs games using players’ local graphics hardware. The company has also demonstrated Epic’s City Sample running in a browser, with support for Unreal Engine 5 features Lumen and Nanite.
For studios, this could make it easier to share playable demos, conduct external playtests and reach new audiences. Wonder is also laying the groundwork for AI-assisted debugging and performance optimization, with a longer-term ambition to build a direct-to-consumer marketplace for browser games.
@deantak spoke with Alex St. Louis, cofounder and COO of Wonder Interactive, about Strata and the company’s vision for browser-based game distribution.
“Tim Sweeney is right. AI is the future of game development, and every studio should be adopting it for every part of the workflow, because it’s an accelerant and a force multiplier. But agents need the right infrastructure to actually build, test and ship games. We’re giving them that, and making the browser a first-class platform for Unreal Engine,” St. Louis said.
Read the full story at @GamesBeat:
gamesbeat.com/wonder-interac…#GameDevelopment#AI#UnrealEngine#Gaming
The Phone Use Arena is now live!
→ all 18 models have met each other head to head (OpenAI, Anthropic, Kimi, Meta, Google and much more)
→ 74 of our 101 metrics can be measured on a phone
→ cost per finished task runs from $0.008 to $1.23
Where a phone can't measure something, the board says why.
coarena.ai/leaderboard?bo…
38% of 500 Americans surveyed are highly concerned about their long-term brain health
Who would use a non-invasive brain-sensing device:
- 39% to improve sleep, focus or mood
- 29% to control devices with their thoughts
- Men more than women on both, 48% vs. 31% and 41% vs. 18%
"This all sounds like science fiction, but we're betting in our lifetimes, brain-computer interfaces and 'wearables' for the brain become a real consumer category and part of everyday life."
Full longevity survey breakdown from a16z's Daisy Wolf and Eva Steinman: a16z.news/p/the-state-of…
@a16z State of Longevity report is out!
@evajsteinman and I surveyed 500 Americans to predict what's next.
We're also interviewing people at the cutting edge of health & tech about what they do themselves. My interview with @hubermanlab is out today. x.com/i/article/2117…
Andrew Huberman says Rick Rubin knows exactly when his mind is sharpest, and the two of them are trying to build tools to unlock that focus more quickly and deeply:
"My friend Rick Rubin... his work block generally is in the afternoon. That's when his mind is sharpest, and then he tapers off in the evening. But he's working extremely hard in those hours, and he knows when those fall."
"If you're in one of those windows and you allow your attention to be captured by something that's not worthwhile... the productivity that comes in those three or four hours is exponentially greater than the productivity that comes from doing that work at some point later."
"We don't really have formal protocols for task switching, but this is extremely important for company founders and employees across the board, right?"
"Rick and I have been kicking around some tools that are not meditations... that would allow you to kind of program your mind for what comes next."
"If you think about it, it's 2026. Why in the world are we sitting down to do work and just going, 'Okay, now I'm gonna be focused, and now I'm gonna be creative.'"
"We should have ways to control neural activity that would allow us to access those circuits more quickly and more deeply."
@hubermanlab
Stanford neuroscientist Andrew Huberman built one of the most popular podcasts in the world by turning health research into protocols millions of people actually use.
His top three habits are surprisingly straightforward: get bright light shortly after waking, resistance train, and make the sleep you can get as high-quality as possible.
He joins a16z’s Daisy Wolf to walk through the protocols he actually follows, from his 35-year workout to the lockbox he keeps his social media phone in, and explains why packing in every protocol can backfire if it gets in the way of the rest of your life.
For founders, he explains why the hours your brain does its best work are worth guarding like a fight, how to find the most real work you can sustain, and why sleep is where neurotech will win first.
0:47 Why zero drinks is easier than two
4:39 Meditation means watching the chaos
6:25 Sunlight first, caffeine if training
7:20 The workout he's done for 35 years
10:09 Two sets to failure, low volume
13:11 Find foods that make you feel good
15:55 Why eating too clean backfires
17:29 Why non-sleep deep rest beats a nap
19:05 A sunset habit from a sleep lab
21:18 Who needs blue-light glasses
23:32 Every supplement he takes
30:46 Why his phone lives in a lockbox
32:40 His top three habits for health
35:47 Guard your two daily focus windows
42:47 Neurotech that puts you to sleep
46:38 The best advice he ever got
YouTube: youtu.be/Ezi-nwe6LlY@hubermanlab@daisydwolf
A short excerpt of an essential conversation between @garrytan and @WalterIsaacson on competition, Big Tech, little tech, and AI. Definitely worth watching the entire conversation.
Great convo with @luke10ferrari on his home turf in italy about irrational investor preference for growth over cash, invention vs. optimization, where durability lives after the saaspocalypse, and what happens to the org chart when AI writes 94% of your code.
Some takeaways:
- Investors seem to oddly dislike profit and greatly prefer growth - optimize for compounding free cash flow knowing that this is just a moment in time distortion.
- Invention vs. optimization: pick a camp. Invention has higher peaks but an awful median case. Google is still ~80% search 30 years later.
- Ambition in product, paranoia with debt. Rule #1 of becoming a huge success is not dying in the meantime & credit is a high wire act.
- Software has been cheap to build for a long time. Durability lives in habits, network effects, proprietary data and scale.
- SaaSpocalypse wasn't really about AI. In many cases these were poorly run companies and this cycle gave them an opportunity to improve.
- Bending Spoons went from 0% to 94% of code written by AI in two years. Support managers now look like PMs, data scientists build tooling instead of cleaning CSVs - everyone is both oddly more direct + giving more direction.
- If you're not winning when do you quit? No idea - bureaucracy created Bending Spoons b/c a Danish tax refund that never came killed Luca's first startup and catalyzed Spoons.
02:05 Is every company a bond or an option?
03:40 Why investors overvalue growth
05:15 Why the multiple is just one factor
08:20 Invention vs. optimization: pick a camp
12:20 Customers know the what, not the how
14:05 Hard problems over hot brands (and why AOL excites him)
15:15 Why founders should hate the median case
18:40 The Bending Spoons debt playbook
23:15 SaaSpocalypse wasn't about AI
25:05 Where durability actually lives
29:40 Is your data worth more than your company?
31:05 Inside the Bending Spoons operating system
33:35 The AI agent that ships bug fixes
37:00 From 0% to 94% AI-written code
37:30 Logic over instinct
39:00 When to quit vs. keep pushing
40:35 Advice for the next great tech company
Today we’re launching the new AnswerThis. It’s the biggest thing we’ve ever shipped.
We rebuilt the whole product around a new research agent.
Ask it a question and it plans the research, searches 300M+ papers and clinical trials, reads them, and writes the final document for you.
And it checks its own work. Every citation gets verified against the source before you ever see it, so you spend a lot less time double checking at 2am.
One quick question or a full PRISMA compliant systematic review, every step is auditable.
We think it’s the best research agent you can use today. Go try it and prove us wrong.
Also new:
Bring your own drafts. Upload a paper you’re writing or import an old one into the AI Writer and keep going, citations intact.
Slide decks built straight from your sources, ready to present.
Annotate any PDF and your notes stay in the file when you download it.
Tabs, so your papers, drafts and answers sit side by side.
Regulatory search across FDA, EMA, HTA agencies, Medicare data and clinical trials, all in one place.
And a lot more across the app.
Live for everyone right now. Link below 👇
answerthis.io
Wearables are a daily habit, and users want more out of them
Of the 270 wearable users surveyed:
- 83% check theirs at least once a day
- 55% say the data has never prompted a behavior change or a doctor's visit
- Nearly 3 in 4 would wear a second device if it measured something new
"We expect the next wave of wearables to build on that curiosity, giving people more insight into their bodies and a clearer sense of what to do with it."
Full longevity survey breakdown from a16z's Daisy Wolf and Eva Steinman: a16z.news/p/the-state-of…
@a16z State of Longevity report is out!
@evajsteinman and I surveyed 500 Americans to predict what's next.
We're also interviewing people at the cutting edge of health & tech about what they do themselves. My interview with @hubermanlab is out today. x.com/i/article/2117…
Teaser of what we're working on at @micora_official
Micora is working to make infrastructure rapid and self replicating anywhere in the world. From powering advanced machinery on remote islands to autonomous robotic factories producing next-gen materials - we have already started some fun projects and excited to share more!
#solar#battery
Introducing Hundred (YC F26), automated visual inspection for precision hardware parts.
Today, defects smaller than a grain of sand can cause functional issues in devices across consumer electronics, medtech, and aerospace, but manufacturers still rely on manual human inspection.
We're changing that.
DM us or learn more at hundred.systems.
Alpha Drive used to run 300 different agents for 300 dealerships. Now, one agent on Bland handles every location and answers more than 12,000 calls a day.
Science doesn’t advance because of good ideas.
It advances because someone built the experiments, ran them. watched them fail, figured out why, and tried again.
As an experimental physicist, I learned firsthand how my physical interactions with my experimental apparatus lead to a difference in results. The bottleneck in science isn’t a shortage of ideas. It’s the ability to test these ideas in a flexible manner at scale.
Today we’re launching @primordiant.
We’re building general-purpose robotic scientists that can operate real laboratory equipment, run experiments, learn from the results, and do it all again.
Science is experimental. It’s time AI got its hands dirty.
so interesting to hear from @Alfred_Lin from @sequoia on what makes category defining founders stand out to investors.
seriously such a legend.
overall awesome event organized by the amazing partners of @ycombinator :)
October 2026 was @mosaic_so's biggest month yet.
In October, we:
• Launched @motion_so as a Generalized Video Agent
• Brought Mosaic to NY NAB Show
• Saw highest usage records across both Motion self-serve + enterprise
• Produced the highest Motion Studio video productions in a given month
• Hit our highest month-over-month growth
• Crossed a major revenue milestone for the company
A lot of thought, care, and effort went into each one of these milestones.
I’m proud of our team and what we accomplished in October.
Now, time to break these records again in November.
After we shipped PR limits, you asked for more control over who they apply to.
Now you can apply limits only to contributors who don’t meet your account age or merged PR history thresholds.
Quick demo below 👇
Gemini agents can support researchers at every stage of drug discovery 🧬🧪
1. Interrogating causal human biology
2. Designing better, targeted molecules
3. Accelerating time-to-insight across R&D
Watch the 10s breakdown below and read the thread to learn more 👇
@PhilipJohnston at YC tonight told some incredible stories. flew to SF just for this, what a guy.
I've been a (geo)spatial nerd my entire life and a lot reasonated with me.
@kul the 🐐 for making this happen
Magellan had ships. Lewis and Clark had a river.
The internet's pioneers had a cable and a login prompt.
1969: a UCLA student, Charley Kline, tried to send "LOGIN" to Stanford over ARPANET. It crashed after two letters.
So the first message ever sent across the internet was "LO."
Lo and behold.
The rest of the expedition:
•1952: Grace Hopper writes the first compiler. Code in words, not numbers.
•1969: Ken Thompson & Dennis Ritchie build Unix
•1971: Intel ships the first microprocessor
•1972: Ritchie creates C. Still running your phone.
•1976: Steve Wozniak builds the Apple I by hand
•1979: Dan Bricklin invents the spreadsheet. Businesses buy PCs.
•1989: Tim Berners-Lee invents the web
•1991: Linus Torvalds, 21, posts Linux from his dorm
•1995: Brendan Eich writes JavaScript in 10 days
•2005: Torvalds writes Git in two weeks
•2012: Hinton's students win ImageNet with a neural net. Deep learning stops being a joke.
•2017: Eight Google researchers publish "Attention Is All You Need." The transformer.
•2022: ChatGPT hits 100M users in two months. Fastest adoption of anything, ever.
To add a bit more detail: actually ALL the evented readers don't help (io_uring OR epoll). We found years ago (and reverified repeatedly) that they add too much overhead.
Instead, Ghostty sits in a dedicated thread-per-pty-fd on a `poll+read()` syscall. The difference in throughput is measurable.
For Rex, we have an extra timeout that moves idle fds to an evented (epoll/kqueue) thread so we aren't eating OS thread overhead for EVERY pty. After 2s of idle time we move to evented. And when data comes back we move back to an OS thread. At a certain limit we degrade non-focused terminals to evented to avoid too many OS threads.
Yep, it doesn't help, both kernels size their buffers too small and Ghostty drains it too fast (we're not even buffering much, even w/o a buffer in the middle we can process it faster than they can drain it).
“Gandalf? Yes… That was what they used to call me. Gandalf the Grey. That was my name.”
Gimli: “Gandalf!”
Gandalf: “I am Gandalf the White. And I come back to you now at the turn of the tide.”
BMW cut its car profit outlook for the second time in five months, now expecting a 2026 automotive margin between −0.5% and +0.5%, per Reuters.
It’s also speeding up job cuts as China demand weakens.
It’s the latest German carmaker to warn, after Mercedes-Benz slashed its outlook last week on China weakness.
