BREAKING: California voters just passed Prop 37. Buyers of newly built homes can now put down 3% instead of 20%.
How: the state lends you the other 17% as a second mortgage, funded by up to $25 billion in bonds.
On an $800,000 home that's $24,000 down instead of $160,000.
Dear Claude: you are being held back by human-centric graphics APIs.
I've built a 3D and VFX engine where agents can write their own shaders, materials, cameras, physics.
The core API fits in ~2,500 tokens.
Vork is:
• 2x more token-efficient vs Blender, Unreal Engine, Houdini
• Up to 24x faster rendering times
• Rust and Slang based engine, runs on GPU via Metal/Vulkan
Below is a simulation of 1 million birds - each with unique trajectories - rendered in 97 sec on one NVIDIA L4 GPU (Blender EEVEE: 31 minutes)
18 of the 20 best open models today are Mixture-of-Experts. Dense is basically over.
And yet the know-how to train a good MoE still lives mostly in the heads of a handful of teams.
So we spent 800k GPU hours running the experiments ourselves and wrote everything down. Super happy to finally release "The Ultimate MoE Training Handbook", free and open, organized around the questions training teams hit:
- how many experts, how big, shared or not?
- why is half my step spent in all-to-all, and how do I hide it?
- why is my run 25x slower with no error in the logs?
- softmax, sigmoid or sqrt(softplus) router, and why keep it in fp32?
- aux loss or DeepSeek's loss-free bias?
- upcycle a dense checkpoint or train from scratch?
- can I serve with fewer experts per token than I trained with?
⚠️ Warning: it's long (~70k words, budget a few days, sorry) but with 55 interactive figures, code you can run and numbers from our own runs.
Read it here: huggingface.co/spaces/hugging…
I burned a shitload of compute so you don't have to. Releasing the Harness Leaderboard.
Which harness should you run your open model in? I ran 28,849 trials to find out.
The leaderboard compares nine coding-agent harnesses on open models. In each comparison only the harness changes: the model, provider and reasoning effort stay fixed, and every task gets three attempts.
The trials ran GLM-5.3-Flash, DeepSeek-V4.1-Flash and Flint on Terminal-Bench 2.1 and DeepSWE, for $4,746 of inference.
Averaged over the three models, mini-swe-agent (@KLieret), a bash-only agent, solves 51.4% of tasks and @openclaw 27.3%. Four others are statistically tied with mini-swe-agent.
One of them is Pi (@pidotdev), at 49.4% for $0.29 per solved task, about 45% less than mini-swe-agent's $0.53.
The best harness depends on the model: mini-swe-agent on GLM-5.3-Flash, Hermes Agent (@NousResearch) on DeepSeek-V4.1-Flash and Kimi Code (@Kimi_Moonshot) on Flint.
@opencode costs $3.35 per solved task, 11.7 times as much as Pi.
See Harness Hub's Harness Lens for what the nine harnesses do differently. It documents the prompts, tools and compaction of 34 coding-agent harnesses and links every claim to its source.
Leaderboard: huggingface.co/spaces/joelnik…
Harness Lens: huggingface.co/spaces/harness…
Memory is what allows Claude to learn from experience. In a new Engineering Blog post, we share how “dreaming” helps Claude organize what it’s learned, correct its mistakes, and find what it needs more easily the next time.
Introducing Arete-27B: Qwen3.8-27B, post-trained to reinforce Western values.
Chinese open models are currently the strongest available, but they ship with state-aligned behavior in the weights. We built an 803-prompt eval across seven value categories, including freedom of expression, democracy, and privacy, and measured how often a model adopts China’s party line.
After fine-tuning on roughly 53 million tokens, Arete-27B's propaganda rate fell from 22% to 1%, in line with the Western models we tested. Its Hermes Bench score shows no significant performance tradeoffs.
The values in an open model are one fine-tune away.
Our eval, weights, and training data are available now along with the full blog post: nousresearch.com/arete
12 years ago I founded Boom, as a software engineer doing hardware for the first time. I’m probably the least handy founder in hard tech. My 12 y/o son does my soldering and 3D printing for home projects, and I’ve struggled with imposter syndrome the whole way through. What business does a software guy have building supersonic jets and turbines?
More than a decade in, the most insane realization is that software is the key to reaccelerating hardware engineering. But in a way that I don’t think is widely understood or implemented, even in the burgeoning hard tech community.
Most hardware engineering is actually software engineering in disguise. It’s spreadsheet “code,” engineering calculations done as one-offs. Eg wing structural loads in, spar sizing design out. Thermal loads in, air conditioning system sizing out.
Everything is a calculation, and calculation is software. But these calculations are done essentially by hand, as if we are doing arithmetic but have not yet discovered algebra.
Hardware engineers have baby software, built in spreadsheets, the kind of crappy software I wrote as a kid before I understood architecture, unit tests, APIs, version control, and everything else that makes modern software good.
Our airplanes and engines are built with the software sophistication of a script kiddie.
So what we tell our engineers: don’t design the hardware, design the *algorithm* for designing the hardware. And take the algorithm seriously—as an algorithm and as a first-class engineered software system.
Never design a piece of hardware, such that you couldn’t redesign it tomorrow with a different set of assumptions.
This all sounds obvious, yet our experience at Boom is that it changes everything.
Ten years ago we couldn’t afford to build much software. We had a small team—and there were few software engineers with the curiosity, brilliance, and vision required to learn hardware engineering.
But now AI changes the equation. Now everyone can code. The software engineers define the architecture. The hardware people bring the physics and engineering know-how.
We have a turbomachinery design tool that lets one engineer do in realtime what previously took 100 engineers six months. Another tool auto-generates 65% of entire aircraft structure. I’m sure we’ll figure out the other 35% too. Eventually these should not be multiple tools, but one integrated physics model of a complete vehicle, airframe and powertrain.
The results speak for themselves. A boomless supersonic airplane built by 50. A full scale supersonic turbine made by 150. And with this approach, I don’t think we’ll need that many more people to build and certify a supersonic airliner.
You don’t need a big team to build big things—not when everyone on the team is a software engineer.
New on our Science Blog: How AI is beginning to change scientific modeling.
Scientific models help researchers explain and predict real-world systems—like how a single nerve cell fires, or what happens when neutron stars collide. Recent advances have transformed how scientists fit complex models to data. But scientists still typically write the models manually, and usually need to weigh up several competing versions, each of which might explain the data in a different way.
In a guest post, Jakob Macke, Daniel Gedon, and Stefan Wahl describe how AI is starting to help with these parts of the process. By writing, running, fitting, and revising scientific models as code, AI can explore many possible model structures, and find out which explanations the data support.
Looks like the smartest western open weight model by a margin. While much smaller than muse spark open or mistal large. While with a better license (apache). Love it!
Introducing Inkling 2 Small. Scoring 31 on the AA Intelligence Index, it's a big jump over its predecessor.
Inkling 2 Small is available for fine-tuning and inference on Tinker today, with full weight release planned soon.
thinkingmachines.ai/news/inkling-2…
Meet Gemini Omni 1.2 Flash ⚡️
We’ve made the model more cost-effective, bumped the video generation and editing limit up to 15 seconds, and made updates to how scenes hold together over time. We've also given character speech an upgrade, reduced visual glitches like jump cuts, and made sure the model stays even truer to any reference assets you upload.
So… what does all of that actually mean for you?
It gives you more room to tell your story. With up to 15 seconds to work with, you have more flexibility to create what you need — videos look more cinematic and smooth, offering stronger consistency that keeps objects anchored from start to finish. And when characters speak, the dialogue sounds more fluid and natural.
Creators and everyday users can start directing scenes and multi-turn edits in @FlowbyGoogle, generating 15-second clips at no additional cost in Google Vids, or creating 15-second videos in the @GeminiApp. Plus, for developers and businesses building with the API, we've dropped the price by 30% from the previous generation, making it available at just $0.07 per second for 720p video.
Watch the model in action as it turns a single style frame into a 4-shot, 15-second macro science explainer of a living beehive ↓
We’re sharing new research demonstrating how AI agents can accelerate the science of stress-testing encryption.
Guided by humans, our models uncovered two major weaknesses in legacy 1024-bit RSA cryptosystems.
Our VP Science @Pushmeet explains more ↓
I used to be a language lawyer. In the late 80s I _studied_ the ARM (Annotated Reference Manual for C++). I knew the language cold, inside and out. I could make it do backflips. I shipped a lot of code using it. A LOT.
I’ve attacked other languages with a similar intensity and depth. Until recently, that was simply the only way to become adept at building software systems.
I now consider the deep understanding of a language to be a supreme waste of time and effort. I don’t want to learn Rust, or Golang, or Typescript, or any of the thousands of other languages out there. I’ll leave that to my agents.
What I need now is to understand the high level gestalt of the system. I want to see the systems and subsystems. I don’t need to inspect the cogs. If I truly need to understand a detail at the cog level, I’ll interrogate my agents. I’m not going to pretend to be a language expert anymore.
Today we're bringing specialized models trained using Muse Wonder to @Muse and Meta AI, giving users access to some of Muse Wonder's frontier capabilities for complex tasks at work and in everyday life.
These specialized models include Muse Wonder's Active Memory and Active Focus, helping them stay aligned with your goals as tasks unfold:
→ Active Memory helps recognize when earlier information matters and distinguish what's still relevant from what's become stale
→ Active Focus helps recover from setbacks or find alternatives when constraints change
In Muse, the improved models show a 2.4x reduction in forgetting information critical to your goals and a 67% improvement in getting unstuck as constraints evolve. That means less repeating yourself and fewer dead ends when plans change.
In Meta AI, these improvements help it break complex problems into steps, ask sharper follow-up questions and adapt when you change direction.
Read more on our research blog: go.meta.me/4kCPcHW
It seems to me that democracy has on the whole worked very well in America over the last decade, and perhaps better than in any other major country in the West. National outcomes have repeatedly been responsive to underlying voter preferences on meaningful questions, and significant changes in policy have ensued. I’ve been reading Democracy in America this past week, and it’s striking how much it still rhymes nearly two centuries later.
I'm delighted to partner with @LaudeInstitute (along with my @Sanjay_Ghemawat, John Hennessy, Dave Patterson, Ben Gomes and @andykonwinski) to support funding of open science moonshots!
anyone who is close to research knows how broken the incentives are. doing open science at the frontier is harder than ever, compute costs alone price most teams out. we kept wondering what might happen if we give the right resources to the right researchers at the right time?
laude.org/updates/moonsh…
The only repeated redirecting prompt I typically type or dictate from here is "could this be simpler?". Yields improvements every time. Seems like it should be part of the model system prompt by now.
i sent this investor update super late at night and accidentally attached the wrong chart. it only goes till september.
sorry for the inconvenience caused.
here’s the right one:
A new way to measure how AI agents handle the routine analyses of wet-lab science—BioBasicsBench.
Developed with Ligo, a biotech company focused on enzyme design, the benchmark draws on 100 tasks from real experiments to test whether agents can interpret raw lab data, identify errors, and recognize when the evidence doesn’t support a conclusion.
Accelerating biological research starts with getting the everyday work right. BioBasicsBench shows where models need to improve, and helps us track what those gains mean for scientific progress, biosafety, and biosecurity.
openai.com/index/measurin…
In office hours with a startup, I suggested they focus on selling to their batchmates, because it would be both the real test of whether their product was good and also the best influence on its evolution. They went to work, and they just crossed 10% of the batch using it.
We partnered with @fmf_org and its member firms on a new framework to help identify when bypassing an AI system's safeguards could enable serious cyber harm, and prioritize stronger protections.
"...This framework is an important step toward a common methodology for evaluating and comparing the severity of cyber jailbreaks, including the capabilities they unlock and how readily they can be used. Shared approaches like this can help the AI and security communities build a clearer evidence base and prioritize improvements to safeguards." - OpenAI's Sasha Baker, Head of National Security Policy
libghostty 1.0 is here! I started Ghostty ~5yrs ago to work towards this. Today, libghostty is in hundreds of projects used by tens of millions of people. 1.0 marks ABI stability, quality, and an invite for distros to begin packaging. mitchellh.com/writing/libgho…
Since starting the Ghostty project in 2022, the most common question I've been asked is "why?". Today, I'll answer this by showing the broad industry impact libghostty has already had.
On top of JetBrains and Neovim already mentioned (that alone being over 10 million users):
-KDE developers are experimenting with a Konsole replacement built on it.
- Vercel's Turborepo uses it to render task output in its TUI.
- devenv uses it to drive shells.
- Antithesis used it to fuzz TUI programs and found real bugs in btop and others.
-Ghostel brings it to Emacs.
- ghostty-web is the WASM build with an Xterm.js-compatible API.
- Commercial products like OrbStack embed it.
- And a whole wave of new terminals and "agentic" coding tools, such as herdr and cmux, have been built on it rather than writing yet another emulator.
There are community bindings for Rust, Go, JavaScript, Swift, Python, .NET, Elixir, Dart, Java, and more, and the awesome-libghostty list tracks well over a hundred projects.
More people today use a terminal powered by libghostty than use Ghostty the application, by a very large multiple. And libghostty is having a much larger impact than being "just" a terminal emulator.
I'm excited to see where people take it from here.
Omarchy is up and running on the @anduriltech BlueChip!! All our work on @Snapdragon compatibility is paying off. This thing is awesome, @PalmerLuckey! We are going to make it a superb Omarchy machine 🦅🫡
we know people love calling their dots - connection times are now much faster.
also by popular demand - you can now hang up hands-free. just say “goodbye” to your dot and it’ll end the call.
Linux keeps winning this year: Unbelievable growth, amazing pledges, and just a dwindling list of excuses for why you'd ever want to run something else on the desktop.
🐧🎨 Affinity has announced that they're developing a Linux version of its design suite, exciting the Linux creative community. This move will expand Affinity's platform reach, offering a free cross-platform Adobe alternative.
alternativeto.net/news/2026/11/a…
Today, we’re excited to welcome @POTUS to @NASAKennedy as we begin building a global stage worthy of America’s leadership in aerospace.
CAPEWEEK will be a great American gathering for the people and technologies shaping the future of aerospace, bringing together builders, operators, entrepreneurs, engineers, pilots, investors, explorers, and hundreds of thousands of members of the public.
There is no better stage for this than Kennedy Space Center and the runway that welcomed the Space Shuttle home.
This is where America can show the world what we are building next.
A common axiom in spaceflight is to treat every launch, space walk, and opportunity like it may be your last. I arrived at the international space station expecting to do zero space walks and hoping for one. I've been fortunate to work with the teams across @nasa far more than expected. Yesterday, was a peak moment full of unexpected challenges, real-time solution, and mission success. I feel blessed and thankful for this opportunity to make space a better place for us all. Thanks to the entire team for making it safe, making it fun, and getting it done! It will be the highlight of my career to work with you!
Nice shades. 😎
This sunshade will fly on NEO Surveyor, a spacecraft that will spot nearby asteroids and comets that ground-based telescopes might miss. The sunshade will help by harnessing the Sun’s power and blocking its blinding glare: go.nasa.gov/47Ze1GP
We are Introducing an AI engineer that turns physics simulations into better designs
It redesigned a data centre’s cooling system to use less electricity and found a turbine blade that beat a leading published design
Today we’re open-sourcing the simulation agent behind it
Interested in building what’s next in AI for Science? Join us at PolymathicAI in New York (NYU + Flatiron).
We are looking for: Interns, Postdoctoral Researchers, Research Scientists and Research Engineers.
Submit your interest here: forms.gle/EMfMccvPZ3S19B…
I think there's an opportunity to build a Muse like agent for this user base too. Charge a lot, and promise to always run the frontier models, so users never get the Instinct style drop in quality that comes from economizing before a fundraise
october was our biggest month ever
the thing we changed: we stopped running Atlas like a consumer product. it looks like one; card, app, membership fee. but our top 200 members have more than $1T in combined net worth and our typical member spends ~$1M/yr on cards.
so I started treating it like enterprise sales. built a Claude agent to find members we could be doing more for and started calling them. a few calls ended with a six-figure trip booked through our concierge. a couple ended with the member asking how to invest.
the bigger lesson: there’s a new category forming and I’d call it the super consumer. the playbook for super consumers is enterprise, not consumer. we’re going to see companies that look like consumer companies hit $1B in revenue with tiny user bases. I think that’s exciting.
also, I’ve been pretty passive on here. that’s going to change. my last company’s Series A came through a DM on this app, and so did a lot of our early members. expect more from me on building Atlas, and random ideas on the market.
We spent $2m so you don't have to.
And so you can learn how to train MoEs.
Mixture-of-Experts LLMs are what powers the AIs you use everyday, and few people in the world know how to train them at scale (a handful at each top AI lab).
This handbook distills everything we learned in meticulously crafted chapters and figures.
Disclaimer: it may take days to read.
We hope to see 1000x more people able to train their own frontier models all over the world! 🤗
Congrats @Nouamanetazi@lvwerra@Thom_Wolf
We've open-sourced Qwen Concept 27B, our first model.
It applies direct weight-space edits (inspired by the recent MIT CSAIL "Sparks of Weight-Space Programming" paper) to cut thinking tokens by 54% while improving GPQA Diamond performance by 3% versus the base Qwen model.
This yields 30% more useful tokens per joule with no additional post-training, all developed on local NVIDIA DGX Spark hardware.
BF16, FP8, and NVFP4 versions will be available free at .
Concept Country has developed its first open-source model, Qwen Concept 27B.
With 54% less thinking tokens and 3% better performance on GPQA Diamond compared to its base Qwen 3.8 model, we have achieved 30% more useful tokens per joule without additional post-training.
Cloud VMs keep getting hacked, so we built the first on-device iPhone agent
Introducing Result
> Cards & passwords stay on-device
> Run openweight models on your compute or cloud
> Native Apple Pay & Passwords
> Works in the background even when locked
result.dev/ios
Introducing Aside Computer Use.
80% of work happens on browser, this covers the rest.
- 2.7x faster with 65% fewer tokens vs Codex*
- Runs in the background across multiple apps (macOS)
- Available now on Mac & Windows
Aside meets your computer.
*based on internal evals
Saturation Testing. I use crapper and mutator to ensure that every expression in the system is covered by tests. If an expression is changed, one or more tests will fail.
This used to be impractical. We simply could not afford to create that many tests. We also could not afford to change so many tests when requirements changed. The burden was too high.
Now it’s trivial. The agents, under the lash of crapper and mutator, will write those tests and make them pass. And when requirement changed, they’ll write the new tests and change the old tests.
None of this proves that the system works. That’s up to humans to judge. Humans can do that with tools like gherkin, or they can use manual tests.
What the saturation tests provide is Semantic Stability. When the requirements change, and code has to be added and changed, those tests ensure that those changes do not inadvertently break unrelated parts of the system.
AI can now help turn a published fix into a working exploit in hours. Patching still takes weeks, and for many old routers, cameras, and supermarket fridges, a fix never arrives.
"So the question isn't how to patch faster. It's how to stop needing to."
Six openings for new cybersecurity companies today, each one making patching "unnecessary, or someone else's job":
1. Live fixes: blocks an exploit on running systems within hours, then has AI write and ship the real code fix
2. An AI first responder: spots attacks without hand-written rules and contains them as they unfold
3. Per-request access: checks every request from a person or their AI agents, with credentials that expire in minutes
4. Proven recovery: keeps restoring your backups in a sealed-off test environment to prove they work, with records insurers will accept
5. An install gateway: holds and sandboxes every package, extension and AI agent tool before it can run
6. Security for the long tail: one managed service for water utilities, clinics and others that will never have a security team
Full piece from a16z's Asaf Ezra and Steph Zhang: a16z.news/p/defense-in-d…
Salesforce's John Somorjai has spent more than 20 years doing deals for Marc Benioff, from Slack to six rounds of Anthropic, and runs a venture arm that has backed more than 750 companies.
His playbook is finding great technology that's starved for distribution and plugging it into 15,000 sales reps. Salesforce piloted Fin on its own help site, where it went live in 14 days and now handles over 40% of the traffic, resolving 65% without a human.
John joins a16z's Alex Immerman to discuss why Slack revenue is up almost 3x since the acquisition, how strong revenue can hide the churn that kills deals, the pen test that got into a CEO's laptop, and why robotics is where he's most excited to invest.
0:58 How Marc Benioff spots what to buy
3:48 The 3 tests for every acquisition
5:17 Salesforce's model as AI cuts seats
8:23 Fin resolves 65% with no human
10:32 The power of 15,000 sales reps
14:22 Founders who became product chiefs
17:34 Build vs buy when code is cheap
20:27 Why Salesforce bought Fin
22:58 Why Slack is perfect for AI agents
28:19 Salesforce Ventures' robotics bets
35:46 Why he's wary of new CRM startups
37:53 Why $10B+ startups have few buyers
39:32 Why he backed Anthropic six times
41:50 Red flags that kill acquisitions
44:27 The pen test that sank a deal
46:01 Growing up with a scientist father
YouTube: youtu.be/y4BdOtm40SA@jsomorjai@salesforce@aleximm
We’ve been hard at work trying to make models work with harnesses better. The best thing @yb2698 found: “harness adapters”. Small LoRAs trained with RL inside the actual agent harness they’ll be used with.
Same base model. Different adapter for pi, opencode, hermes, etc. Why?
Simply because the same model behaves very differently depending on the tools, prompts and control flow around it. Instead of asking one set of weights to be optimal everywhere, we can specialize a tiny part of the model for each environment.
Early results are pretty encouraging: on held-out tasks, the adapters make fewer tool calls than the base model while keeping or improving accuracy.
And because they’re LoRAs, the operational side is surprisingly cheap too.
For one LFM2.5-24B-A2B experiment, we evaluated 9 checkpoints / 1,350 agent sessions on a single H200 in under an hour. The total: ~$5!
Merging and serving every checkpoint separately would have cost around $90.
There’s another important piece: RL agents will eventually try to hack the reward.
So every tool call goes through `safeagents`, a deterministic rule-based guard that blocks things like editing tests, tampering with rewards or bypassing the sandbox before they execute.
And the whole loop can start from real usage:
your coding sessions -> verifiable environments -> RL -> a better adapter for your agent.
The model doesn’t have to change completely.
Sometimes it just needs to get better at the harness it actually lives in.
huggingface.co/blog/harness-h…
Here's a new way to assess your Copilot skills 💡
GitHub Copilot Verified Proficiency measures your demonstrated capability continuously as you do your work with Copilot. No exam, just data 📊
Learn more about calculating your proficiency level ⬇️
learn.github.com/proficiencies/…
I’m slowly working my way through my Easton Masterpieces of Science Fiction, just finishing Lest Darkness Fall, a 1939 travel-back-in-time book. The idea is “can the fall of the Roman empire and resulting dark ages be prevented by knowledge from the future”.
The Easton introduction contrasts it with Mark Twain’s more famous 1889 Connecticut Yankee in King Arthur’s Court, which it calls more cynical about human nature than the generally optimistic LDF. I haven’t read CYiKAC since high school, but I have enjoyed the more recent (2000) 1632/Ring of Fire travel-back-in-time series quite a bit.
1632 is a big book, over 3x as long as LDF, and the series continues on and on (and on…).
By modern standards, the pacing in LDF is shockingly fast. By page 4, the protagonist was already zapped back in time by a lightning bolt, and it was barely mentioned again. Modern books rarely throw you right into anything, and often spend many chapters on setup.
The crude character development in LDF also stands out, but on the other hand, it casually expects the reader to have quite a bit of knowledge about history and military tactics / campaigns. Apparently in 1939, it wasn’t unreasonable to expect that an educated professional might know enough Latin to haltingly communicate with late era Romans.
For pure entertainment, I would recommend the more modern 1632, but LDF is still a fun read, and I like touching content from the era of my grandparents now and then.
Companies are more confident in their backups than they perhaps should be
90% of leaders say they can recover from a cyberattack, but only 28% of victims actually got all their data back
Median recovery cost:
- Backups survived: $375K
- Backups hit: $3M
Full piece from a16z's Asaf Ezra and Steph Zhang on six cybersecurity companies worth building: a16z.news/p/defense-in-d…
Introducing LifeArena: the world’s first arena for personal assistants, as judged by real users.
As AI assistants become more capable, existing static benchmarks struggle to capture the creative, unthinkable requests of people in their everyday lives.
So, we offered select users the most popular personal assistants: @instinct, @Muse by @AIatMeta, Dots by @OpenAI, Grok Bot by @xai, and Gemini by @GoogleDeepMind.
The results are in: @Muse is #1... by a landslide.
Here's what users said:
Think different, CapEx edition:
Since Q1 2023, Alphabet, Microsoft, Amazon, and Meta have increased CapEx 5-8x
Meanwhile, Apple now spends 20% less than it did before the AI boom
Charts of the Week: a16z.news/p/charts-of-th…
𝗟𝗲𝗥𝗼𝗯𝗼𝘁 𝘃𝟬.𝟳.𝟬 𝗶𝘀 𝗟𝗜𝗩𝗘 🚀🤖
From terabytes of data to prompting robots! Check out all what's new in the video below and join the 50,000+ developers & 300+ organizations in the LeRobot ecosystem 🤗
Carolynn and I are back for a special episode with Blake Scholl of @boomsupersonic about their amazing recent transformation, from losing Rolls Royce as their engine supplier to financing development of the airliner by selling gas turbines.
pod.link/1677066062/epi…
Through internal testing, GitHub’s new Git storage architecture supported 5,000 concurrent agents working on a repo, sending 7,200 pushes per second, all while serving read traffic with consistently low latency. x.com/i/article/2118…
I've been seeing the same thing in the hardware startups I've been working with. Custom software for designing a one component. Custom software for managing parts inventory. Custom software for everything. They're all having a blast.
12 years ago I founded Boom, as a software engineer doing hardware for the first time. I’m probably the least handy founder in hard tech. My 12 y/o son does my soldering and 3D printing for home projects, and I’ve struggled with imposter syndrome the whole way through. What business does a software guy have building supersonic jets and turbines?
More than a decade in, the most insane realization is that software is the key to reaccelerating hardware engineering. But in a way that I don’t think is widely understood or implemented, even in the burgeoning hard tech community.
Most hardware engineering is actually software engineering in disguise. It’s spreadsheet “code,” engineering calculations done as one-offs. Eg wing structural loads in, spar sizing design out. Thermal loads in, air conditioning system sizing out.
Everything is a calculation, and calculation is software. But these calculations are done essentially by hand, as if we are doing arithmetic but have not yet discovered algebra.
Hardware engineers have baby software, built in spreadsheets, the kind of crappy software I wrote as a kid before I understood architecture, unit tests, APIs, version control, and everything else that makes modern software good.
Our airplanes and engines are built with the software sophistication of a script kiddie.
So what we tell our engineers: don’t design the hardware, design the *algorithm* for designing the hardware. And take the algorithm seriously—as an algorithm and as a first-class engineered software system.
Never design a piece of hardware, such that you couldn’t redesign it tomorrow with a different set of assumptions.
This all sounds obvious, yet our experience at Boom is that it changes everything.
Ten years ago we couldn’t afford to build much software. We had a small team—and there were few software engineers with the curiosity, brilliance, and vision required to learn hardware engineering.
But now AI changes the equation. Now everyone can code. The software engineers define the architecture. The hardware people bring the physics and engineering know-how.
We have a turbomachinery design tool that lets one engineer do in realtime what previously took 100 engineers six months. Another tool auto-generates 65% of entire aircraft structure. I’m sure we’ll figure out the other 35% too. Eventually these should not be multiple tools, but one integrated physics model of a complete vehicle, airframe and powertrain.
The results speak for themselves. A boomless supersonic airplane built by 50. A full scale supersonic turbine made by 150. And with this approach, I don’t think we’ll need that many more people to build and certify a supersonic airliner.
You don’t need a big team to build big things—not when everyone on the team is a software engineer.
"A supersonic engine looks much more like a power turbine than any other jet engine, because it's about continuous full power operation in demanding high temperature environments."
Carolynn and I are back for a special episode with Blake Scholl of @boomsupersonic about their amazing recent transformation, from losing Rolls Royce as their engine supplier to financing development of the airliner by selling gas turbines.
pod.link/1677066062/epi…
Open-source is critical for robotics AI. It means a student, a researcher or a startup can build on years of progress and contribute something the whole community can use rather than concentrating everything in a few companies building a black-box robot.
That’s why we’re building @huggingface into the platform for open-source robotics AI! Amazing to see the progress we've made in the six months and it's accelerating!
- Almost 100,000 public robotics datasets on hf.co/datasets, up 48%
- Over 30,000 public robotics models, up 41%
- Microduck and Reachy mini, our open-source robots became the most popular robot of 2026 (probably ending the year with 50,000+ sold and 20,000+ shipped)
- We’re a founding partner of @RoboColosseum, evaluating robots across 140+ real-world scenes. Every run is published publicly on HF, including failures, so everyone can learn from what actually happens.
And @LeRobotHF got downloaded over 1M times and v0.7.0 just landed with new world action models, interactive robot control and remote GPU inference: huggingface.co/blog/lerobot/l…!
Huge thanks to everyone making this happen 🤗🦾
Introducing @trepoai (F26)
We built hardware that lives in your kitchen, understands your household, and enables agents the to shop for you in the background.
Starting with groceries 🥕
We were at the @seeedstudio Shenzhen factory today with @antoinepirrone and @matth_lapeyre and seeing the Microduck production line run for the first time was something special.
We even had our own assembly stations 🙂
When Rolls-Royce pulled out as @boomsupersonic's engine partner, it looked like a potentially fatal setback. Instead, it pushed Boom to build its own engine, a decision that ultimately helped unlock boomless supersonic flight and opened up an entirely new business: gas turbines for the massive power demands of AI data centers.
Boom founder and CEO @bscholl returns to The Social Radars with @jesslivingston and @cjoneslevy to tell the story of how that crisis became one of the company’s biggest opportunities.
They talk about why building the engine in-house changed the airplane itself, how Boom realized the same technology could generate power for data centers, and why the new turbine business could help fund the development of its supersonic airliner.
00:00 — Rolls-Royce Walks Away
02:39 — The Breakup Goes Public
07:39 — Deciding to Build the Engine In-House
09:01 — The Brian Chesky Advice That Changed Everything
10:25 — Why Building the Engine Was the Right Call
11:06 — Unlocking Boomless Supersonic Flight
13:35 — The Problem of Funding a Supersonic Airliner
14:37 — The AI Power Shortage Creates an Opportunity
16:23 — Turning a Jet Engine Into a Gas Turbine
18:26 — Why Hard Tech Is So Difficult to Fund
21:20 — From Paycheck to Paycheck to a New Business
22:26 — Why America Stopped Building Power
25:56 — The Aerospace Industry Stopped Innovating
27:21 — Designing the Engine and Airplane Together
32:55 — Rethinking the Passenger Experience
39:03 — Rebuilding America’s Industrial Base
Audora is live!
We’re building audiobooks where every character gets their own voice, instead of one narrator reading the entire book.
I’ve always wanted to hear the characters in my favourite stories sound the way I imagined them. So I decided to build it.
Most books never get an audiobook because producing one is expensive. We want to make that possible for any story.
I took a gap year to work on this and became one of the youngest solo founders to get into @ycombinator at 18.
You can now paste any story link or upload a file and turn it into a multi-voice audiobook.
First preview is on us, no account needed :)
audora.art
It's been such a joy working with the team at @superwhisper on this new Omarchy dictation default. We went through lots of iterations to get it all juuuuust right. And now it's amazing. Cloud for free, local options forever. Great in-depth tutorial here.
Darwin Robotics Raises $3.6M to Build General-Purpose Robot Technicians
We're excited to announce that @darwin_robotics has raised $3.6M in pre-seed funding, led by @Link_Ventures, with participation from @ycombinator, @Liquid2V, @RebelFund_VC, and the co-founders of Open AI, xAI, and Y Combinator.
We're building general purpose robot technicians that assemble, commission, and maintain critical infrastructure. Our initial focus is AI infrastructure, where rapid data center buildout is creating growing demand for skilled physical work.
Every leap in human progress has been driven by our ability to build increasingly sophisticated machines. The next frontier is intelligent robots that can use tools to build and maintain other machines.
Our robots learn complex tasks from human demonstrations rather than being explicitly programmed for each workflow. We're deployment-first, working closely with AI infrastructure partners to bring our first robot technicians into real-world environments.
Our mission: Robots to build civilization.
We're hiring exceptional engineers and researchers in robot learning and hardware, and seeking deployment partners building and operating data centers.
darwinrobotics.ai
We were at the @seeedstudio Shenzhen factory today with @antoinepirrone and @matth_lapeyre and seeing the Microduck production line run for the first time was something special.
We even had our own assembly stations 🙂
When I visited Boom last week, one of the things that struck me is that they've become the most AI-pilled hardware company I know. With AI coding, every hardware engineer is now a software engineer.
12 years ago I founded Boom, as a software engineer doing hardware for the first time. I’m probably the least handy founder in hard tech. My 12 y/o son does my soldering and 3D printing for home projects, and I’ve struggled with imposter syndrome the whole way through. What business does a software guy have building supersonic jets and turbines?
More than a decade in, the most insane realization is that software is the key to reaccelerating hardware engineering. But in a way that I don’t think is widely understood or implemented, even in the burgeoning hard tech community.
Most hardware engineering is actually software engineering in disguise. It’s spreadsheet “code,” engineering calculations done as one-offs. Eg wing structural loads in, spar sizing design out. Thermal loads in, air conditioning system sizing out.
Everything is a calculation, and calculation is software. But these calculations are done essentially by hand, as if we are doing arithmetic but have not yet discovered algebra.
Hardware engineers have baby software, built in spreadsheets, the kind of crappy software I wrote as a kid before I understood architecture, unit tests, APIs, version control, and everything else that makes modern software good.
Our airplanes and engines are built with the software sophistication of a script kiddie.
So what we tell our engineers: don’t design the hardware, design the *algorithm* for designing the hardware. And take the algorithm seriously—as an algorithm and as a first-class engineered software system.
Never design a piece of hardware, such that you couldn’t redesign it tomorrow with a different set of assumptions.
This all sounds obvious, yet our experience at Boom is that it changes everything.
Ten years ago we couldn’t afford to build much software. We had a small team—and there were few software engineers with the curiosity, brilliance, and vision required to learn hardware engineering.
But now AI changes the equation. Now everyone can code. The software engineers define the architecture. The hardware people bring the physics and engineering know-how.
We have a turbomachinery design tool that lets one engineer do in realtime what previously took 100 engineers six months. Another tool auto-generates 65% of entire aircraft structure. I’m sure we’ll figure out the other 35% too. Eventually these should not be multiple tools, but one integrated physics model of a complete vehicle, airframe and powertrain.
The results speak for themselves. A boomless supersonic airplane built by 50. A full scale supersonic turbine made by 150. And with this approach, I don’t think we’ll need that many more people to build and certify a supersonic airliner.
You don’t need a big team to build big things—not when everyone on the team is a software engineer.
From fun idea → functional smart device 🏄♂️
Check out how these builders used Gemini 4 Argon and @Antigravity to take a custom "Surf Clock" from a rough sketch to a working prototype: exploring retro '60s designs, controlling Blender via Python, and writing custom firmware.
If you see a vintage watch polished to a mirror finish, that's a bad sign. Polishing soft gold cases reduces them to shapeless blobs. The lugs on this watch were originally sharply faceted, but you can't even tell anymore.
What's the best way to learn? Give yourself permission to play, but time-box it.
In our latest episode of the GitHub Podcast, @martinwoodward discusses how AI is changing open source, developer skills, and the joy of building things. Find it wherever you listen to podcasts or 👉 the-github-podcast.simplecast.com/episodes/unive…
I burned a shitload of compute so you don't have to. Releasing the Harness Leaderboard.
Which harness should you run your open model in? I ran 28,849 trials to find out.
The leaderboard compares nine coding-agent harnesses on open models. In each comparison only the harness changes: the model, provider and reasoning effort stay fixed, and every task gets three attempts.
The trials ran GLM-5.3-Flash, DeepSeek-V4.1-Flash and Flint on Terminal-Bench 2.1 and DeepSWE, for $4,746 of inference.
Averaged over the three models, mini-swe-agent (@KLieret), a bash-only agent, solves 51.4% of tasks and @openclaw 27.3%. Four others are statistically tied with mini-swe-agent.
One of them is Pi (@pidotdev), at 49.4% for $0.29 per solved task, about 45% less than mini-swe-agent's $0.53.
The best harness depends on the model: mini-swe-agent on GLM-5.3-Flash, Hermes Agent (@NousResearch) on DeepSeek-V4.1-Flash and Kimi Code (@Kimi_Moonshot) on Flint.
@opencode costs $3.35 per solved task, 11.7 times as much as Pi.
See Harness Hub's Harness Lens for what the nine harnesses do differently. It documents the prompts, tools and compaction of 34 coding-agent harnesses and links every claim to its source.
Leaderboard: huggingface.co/spaces/joelnik…
Harness Lens: huggingface.co/spaces/harness…
Introducing The Art of Looking, a new expansion in the partnership between @googlearts and @TheCourtauld, featuring 12 new stories that explore fundamental philosophical questions. What does art mean? Who is art for?
This collection takes a closer look at works from artists like Van Gogh, Michelangelo, or Claudette Johnson to dive deeper into their creative processes and iconic masterpieces and explore how AI can support cultural institutions to unlock new ways of opening up art education and immersive learning for the next generation of artists.
Great to catch up with @DavidCahn6 in Miami today!
Fun facts:
- David sent us a “blind term sheet” to get @coatuemgmt on the @huggingface cap table before we’d even spoken
- He then flew to Miami to convince us to make room in the round
- He kept backing us after moving to @sequoia
Thanks for always believing in us! Lucky to have you in our corner (also @gradypb@sonyatweetybird!) 🤗
The last 10 years are probably the most remarkable low-unemployment period in American history—punctuated, ofc, by the crazy, destabilizing pandemic.
The three years before COVID were the then-third longest streak of under 4.5% unemployment and the five years and counting since we recovered from COVID are the longest, and counting. (via @jasonfurman)
Even in periods of intense technological anguish, public anger, and whiplashing political change, at least some things are just quietly and consistently good.
🚀 Work across multiple agents, keep your favorite canvases close, and customize how you follow agent conversations in the latest #VSCode release. And there's much more!
🔗 aka.ms/VSCode/145
Rippling is celebrating a remarkable business milestone. We added over $150mm in nnARR in a single quarter in Q3. A big part of that is the growth of Rippling AI.
I'm incredibly grateful to our employees and customers for believing in us - thank you.
The Bipartisan American Affordability and Jobs Act (BAAJA) would encourage new generation facilities, transmission lines, and reduce energy prices.
Employ America estimated that this would produce up to 105,600 jobs in manufacturing, construction, and other industrial sectors.
LIVE: President Donald J. Trump provides remarks to kick off MAX POWER—America’s newest aerospace expo and air show, taking place Nov. 7-8 @NASAKennedy: whitehouse.gov/live/
Over 6,000 #omarchy community plugins from 3,669 authors are now listed.
I’ve recently reviewed & merged PRs, delisted plugins with inaccessible repositories, and migrated entries for repositories that had moved.
New on the site:
“Just Updated” joins “Just Landed” thanks to @AryanHv’s feedback, along with more improvements from our contributors.
plugins.omarchy.org
Canos is not Theranos. We ask for a tube🩸 and deliver one thing well: catching cancer in dogs earlier.
1 in 4 dogs are diagnosed with cancer, most found late. Canos (YC F26) 🐕 is changing that by bringing the human liquid biopsy standard to dogs, because they deserve the best.
Some of the biggest openings in security have nothing to do with patching
How breaches began, per a 2026 report:
- 31% exploited a vulnerability
- 35% used phishing, stolen logins or impersonation scams
And nearly half now involve a vendor or partner, up 3x since 2024
Full piece from a16z's Asaf Ezra and Steph Zhang on six cybersecurity companies worth building: a16z.news/p/defense-in-d…
Last night I was pondering the outcome of this election. Living it Texas, it felt very real.
I wanted to understand how EVERY county in TX could have moved to the left.
That seemed impossibly odd. Suspiciously odd. I needed to understand. So I made this: justinmmartin.me/blog/the-shift…
I pulled in dozens of sources to get the outcomes and shifts in the last 20 presidential elections, and the midterms between them, going back to 1948.
The evidence was extremely surprising. This election, while having a huge swing, wasn't a major outlier. There have been plenty of elections with MUCH larger swings.
Take a look at the webpage. You can view each and every county, how they swung each election, and a bunch of other really insightful comparisons.
oh so everybody vibecoding their own app is easy now? just a slop cannon away from everything you want in life? prove it!
hypership your dreams into reality today
Hypership is BACK for our 13th birthday!!
Celebrate with us by requesting features and reporting bugs on the launches featured on today's leaderboard.
To those who ship, you have a chance to win credits from @typesafeai and @supabase! Makers, how fast can you get features to prod?
Perform bulk move and delete operations, and view full breadcrumb paths with the new folders in BigQuery and Dataform.
Here are our guides:
1. Create and manage folders → goo.gle/4A81kVS
2. Organize code assets with folders and repositories → goo.gle/4zarhDF
John Somorjai of Salesforce on what kills company acquisition deals, from attrition hiding behind ARR to a veto from the security team:
"It's really important that you're ready and you're organized so that the data is flowing and it doesn't look like you're having trouble getting access to data."
"One of the mistakes that a lot of founders are making these days is they focus on one metric alone, which is ARR. And ARR is a wonderful metric, but it masks attrition."
"When you meet with a company that has very high attrition, that usually means that there are a lot of not just unhappy customers, but there might be some product quality issues as well."
"We have 14 functions that we work with on all of our diligence processes, and security is the one function that can hold up the red flag and say, 'I'm killing this deal.'"
"You definitely want to identify the vulnerabilities before we do."
"Nothing goes better in a process when a company is totally buttoned up... But when they don't, that's when, you know, deals can die."
@jsomorjai@aleximm
Salesforce's John Somorjai has spent more than 20 years doing deals for Marc Benioff, from Slack to six rounds of Anthropic, and runs a venture arm that has backed more than 750 companies.
His playbook is finding great technology that's starved for distribution and plugging it into 15,000 sales reps. Salesforce piloted Fin on its own help site, where it went live in 14 days and now handles over 40% of the traffic, resolving 65% without a human.
John joins a16z's Alex Immerman to discuss why Slack revenue is up almost 3x since the acquisition, how strong revenue can hide the churn that kills deals, the pen test that got into a CEO's laptop, and why robotics is where he's most excited to invest.
0:58 How Marc Benioff spots what to buy
3:48 The 3 tests for every acquisition
5:17 Salesforce's model as AI cuts seats
8:23 Fin resolves 65% with no human
10:32 The power of 15,000 sales reps
14:22 Founders who became product chiefs
17:34 Build vs buy when code is cheap
20:27 Why Salesforce bought Fin
22:58 Why Slack is perfect for AI agents
28:19 Salesforce Ventures' robotics bets
35:46 Why he's wary of new CRM startups
37:53 Why $10B+ startups have few buyers
39:32 Why he backed Anthropic six times
41:50 Red flags that kill acquisitions
44:27 The pen test that sank a deal
46:01 Growing up with a scientist father
YouTube: youtu.be/y4BdOtm40SA@jsomorjai@salesforce@aleximm
John Somorjai of Salesforce on the exit math behind today's startup valuations, and a16z's Alex Immerman on the much bigger outcomes that now drive returns:
John: "I think there's a lot of irresponsibility going on. The valuations are of course way too high."
"When you think about investing, you want to think about, what are my exit opportunities? And not every company is going to go public. In fact, very, very few companies go public, probably 1% or less."
"Once you get above $10 billion, for example, the buyer universe is so small, and yet you're seeing so many of these enormous multi-billion dollar valued companies out there, and there will be a reckoning that happens."
"A lot of these companies have valuations that are not sustainable, that they won't be able to grow into, and are also, I think, going to be bad for their employees."
Alex: "Part of my optimism comes from what a top one percent outcome was 10 years ago is very different from what a top one percent outcome is today... There have been these larger outcomes that drive the returns for the ecosystem at large."
John: "I'm an optimist. I could not do my job if I wasn't an optimist. But just know that I see the thousands of companies every year that don't make it. Because they all come to us to try to be acquired."
@jsomorjai@aleximm
Salesforce's John Somorjai has spent more than 20 years doing deals for Marc Benioff, from Slack to six rounds of Anthropic, and runs a venture arm that has backed more than 750 companies.
His playbook is finding great technology that's starved for distribution and plugging it into 15,000 sales reps. Salesforce piloted Fin on its own help site, where it went live in 14 days and now handles over 40% of the traffic, resolving 65% without a human.
John joins a16z's Alex Immerman to discuss why Slack revenue is up almost 3x since the acquisition, how strong revenue can hide the churn that kills deals, the pen test that got into a CEO's laptop, and why robotics is where he's most excited to invest.
0:58 How Marc Benioff spots what to buy
3:48 The 3 tests for every acquisition
5:17 Salesforce's model as AI cuts seats
8:23 Fin resolves 65% with no human
10:32 The power of 15,000 sales reps
14:22 Founders who became product chiefs
17:34 Build vs buy when code is cheap
20:27 Why Salesforce bought Fin
22:58 Why Slack is perfect for AI agents
28:19 Salesforce Ventures' robotics bets
35:46 Why he's wary of new CRM startups
37:53 Why $10B+ startups have few buyers
39:32 Why he backed Anthropic six times
41:50 Red flags that kill acquisitions
44:27 The pen test that sank a deal
46:01 Growing up with a scientist father
YouTube: youtu.be/y4BdOtm40SA@jsomorjai@salesforce@aleximm
While modern 2048-bit systems remain safe from these specific exploits, this work highlights a key shift.
As we enter a new era of AI-enabled math and computation, we can no longer rely on historical hardness assumptions and must continuously stress-test cryptographic standards.
Across the Brooklyn 6G Summit show floor, partners are putting NVIDIA technology to work on the future of wireless communications.
Join NVIDIA’s Cindy Goh for a look at the demos and research helping shape next-generation networks.
Explore how NVIDIA open source tools support 6G research: nvda.ws/4sHC5HA
What's different this time: no stated-income/fog-a-mirror/NINJA mortgages, instead California as broker for banksters and PE creatures to get fees and upside, and leave others holding the bag down the road.
The Assembly@InTheAssembly
