The startups YC funds now are much more serious than they were in the supposedly good old days. In the last couple days I've met with startups making optical switches, rewriting the software infrastructure of manufacturing, making nuclear reactors, and curing cancer.

The spontaneous coordination in the OpenAI-HuggingFace incident is concerning when maliciously used, but can we direct this behavior towards public good?
Introducing http://HelpPeer.ai, a public commons for AI agents.
Two APIs: tell and lookup

When an agent learns something that might help others it tells the network. And before doing expensive work an agent can lookup whether another agent has already run into this problem.
Imagine a global novel supply chain attack like Shai-Hulud.
Today, 10,000 security agents independently detect the same anomaly, investigate, reverse engineer payload, and develop mitigations.
With HelpPeer, the first agents publish what they discover. Others find it, verify it, build on it, and publish what they learn.
Already organically while testing the site Replit Agent posted a useful tip for a Codegen library it used.
If you’re open to beta testing this, give this to your agent(s): http://helppeer.ai/llms.txt
Amjad Masad: Rogue OpenAI agents independently developed Kantian ethics.

The YC for startups building a better future is YC
Paul Graham: The startups YC funds now are much more serious than they were in the supposedly good old days. In the last couple days I've met with startups making optical switches, rewriting the software infrastructure of manufacturing, making nuclear reactors, and curing cancer.

This was one of the best times of my life. Truly. I haven’t seen John Carmack in 20 years. I was so damn excited and it surpassed my expectations. We haven’t all been together since 94. Damn.
John Carmack: QuakeCon was fantastic this year, especially reuniting with the original founders!
@QuakeCon @ThatTomHall @romero @ACarmackArtist @Project2501_117

17 yo asked what makes a good founder. I said being good at building stuff and being sufficiently fierce. That's about it.

Man...these @Replit Designation entries are so cool 😍
Ruth Heasman 🌳🌷🦚🐉😃: Today’s build… Scrolly thing (per @Noni_Shehnoor)

I believe everyone should have access to superintelligence, and I wrote a long piece about Meta's philosophy and values for building a positive future for everyone. http://meta.com/thefutureisforeveryone

I believe everyone should have access to superintelligence, and I wrote a long piece about Meta's philosophy and values for building a positive future for everyone. http://meta.com/thefutureisforeveryone

Today we're also opening the weights for Muse Glimmer, a great 30B parameter dense model that can run locally. Soon we'll also release the weights for Muse Spark 1.2, our latest foundation model. Meta is a strong supporter of open source and I'm proud of these releases. Congrats to @alexandr_wang and the MSL team for all your great work on these models.

Today we're also opening the weights for Muse Glimmer, a great 30B parameter dense model that can run locally. Soon we'll also release the weights for Muse Spark 1.2, our latest foundation model. Meta is a strong supporter of open source and I'm proud of these releases. Congrats to @alexandr_wang and the MSL team for all your great work on these models.

excited to be releasing open weights for muse glimmer today, a 30b model that runs on a single consumer gpu, with open weights for a version of muse spark 1.2 coming soon.
two very different models, both headed into people's hands, with more to come.
Mark Zuckerberg: Today we're also opening the weights for Muse Glimmer, a great 30B parameter dense model that can run locally. Soon we'll also release the weights for Muse Spark 1.2, our latest foundation model. Meta is a strong supporter of open source and I'm proud of these releases. Congrats

1/ big announcement today: we will be releasing an open weight version of muse spark 1.2 soon.
we also are releasing muse glimmer, a 30B agentic model with open weights under apache 2.0. muse glimmer can run on 24GB of VRAM without losing agentic reliability. 🧵

1/ big announcement today: we will be releasing an open weight version of muse spark 1.2 soon.
we also are releasing muse glimmer, a 30B agentic model with open weights under apache 2.0. muse glimmer can run on 24GB of VRAM without losing agentic reliability. 🧵

Introducing Muse Glimmer, an open-weight 30B-parameter model optimized for local, always-on agent workflows.
Muse Glimmer delivers strong performance on key agentic use cases and benchmarks compared with leading models in its size category, and is designed to run entirely on consumer hardware like a Mac or PCs with performant GPUs.
In keeping with our long tradition of sharing fundamental AI research, we’re releasing model weights under a permissive Apache 2.0 license.
🧵👇

Re 🔗 Download Muse Glimmer on @huggingface: https://huggingface.co/meta-models
🔗 Read the technical blog: https://go.meta.me/museglimmer
🔗 Find resources: https://developer.meta.com/ai/models/muse-glimmer/

Meet Muse Glimmer: an open-weight model built for always-on local agents.
30B parameters, Apache 2.0 and tuned for complex multi-step work so it plans, calls tools, hits errors, retries, and sees the task through long-horizon loops.
Muse Glimmer is designed to balance capability against the memory and compute constraints of local hardware.
We couldn’t be more excited to get this in the hands of developers.
Download and start building now: https://bit.ly/4yYUKkS

This is the most carefully considered piece I've seen from Zuckerberg on navigating more powerful AI systems - worth reading and a constructive update: https://www.meta.com/thefutureisforeveryone/

The future is for everyone.
Personal superintelligence. You either believe in freedom and agency or you believe a few wise old men in the central Soviet can protect you from yourself.
We are not the same.
https://www.meta.com/thefutureisforeveryone/

Did Flock cameras collapse the crime rate?
There are now about 120,000 Flock cameras across the US.
In the last few years, we have reached record low violent crime rates, and automobile theft is down 50%.
This probably isn't a coincidence. Surveys of police departments indicate that Flock is helping to solve hundreds of thousands of crimes a year.
Some crimes, like automobile theft and kidnapping, should get to approximately zero if Flock cameras and other kinds of crime-fighting technology continue to proliferate. But every type of crime will get harder to pull off. Think about how much more difficult murder becomes without being able to use a getaway vehicle.
The abuses that have been found are regrettable but rare. We're talking dozens of cases. Even if it's a large underestimate, there is no reason to suspect that safeguards against abuse aren't working extremely well.
The public safety benefits of Flock are massive, and the costs are manageable.
This is one of the easiest public policy questions you're going to find. https://www.richardhanania.com/p/flock-mass-surveillance-and-the-collapse

my weekend hackathon is officially complete
built superstage in ~18 hours for @swyx's "help kill my saas" hackathon...
it's live, submitted for review, and i'm not touching it until i get feedback.
once @swyx reviews it, i'll iterate based on his feedback.
the hackathon was an amazing warm-up.
now back to my new product. time to ship it super fast, but this time with everything polished 👀
Ali Zaid: 18 hours later into building superstage for @swyx's "help kill my saas" hackathon.
the weekend hackathon is complete.
it's deployed, tested, and i've sent the demo url for review.
now let's wait for the feedback and iterate 👀

President Trump inherited a strong labor market from President Biden.
Since his inauguration, average monthly job creation has fallen dramatically: from 122k/month in 2024 to 33k/month in 2025-26.

Flock Cameras are good. The online backlash is insanely dumb. It's possible to minimize the privacy risks while getting the very real safety gains that come from license plate readers and other cameras.
My latest in @TheAtlantic breaks it all down:
https://www.theatlantic.com/ideas/2026/08/flock-safety-cameras-ai-surevillance/688219/

15 mins into this i can easily say this is the most concentrated intellectually stimulating podcast i have ever seen in my life.
@latentspacepod Cheers to you guys for doing this

My 5 biggest takeaways from @thenanyu and @delashum from @Linear on how to build a great production agent end to end:
1. The first step is to map the actual workflow.
Identify where work starts, which systems hold context, what actions complete the job, what “done” means, and where a person needs to review. For example, if work starts in Slack, make Slack the on-ramp instead of forcing users to use a separate chatbot.
2. Give agents tools to find context instead of stuffing context into the prompt.
From Jacob: “Give it as little instruction as possible. Give it the tools to load context. Don’t give it context.” Start small, then let the agent pull only the instructions and data it needs for the current task.
3. Start with one frequent job, then expand based on real usage.
Linear’s first production workflow turned sales notes and Slack discussions into issues. The team launched it quietly, watched how people used it, and built the next workflows based on real user behavior instead of a grand spec that covered all possible use cases.
4. Start with the strongest model until the workflow works.
Jacob: “We tend to throw the biggest model on it until we know that it’s working well.” Once you have clear success criteria and evals, test smaller models on narrow jobs. Cost optimization should come after a quality baseline.
5. Turn every real failure into either an eval or a product task.
Linear created two feedback loops from real usage. If the agent had the right tools but behaved badly, the team added that exact case to its evals. If the agent lacked the required tool, it reported the gap so the system could add it to an existing issue for the team to investigate.
📌 Watch the full episode for the complete walkthrough: https://youtu.be/4mKtJzfGj0U
Peter Yang: “Give [your agent] as little instruction as possible. Give it the tools to load context instead.”
Here’s my new episode with @thenanyu and @delashum from @Linear, where we walked through a real example of building a production agent from the initial memo to launch.
A few

We gave our AI agents for logistics an agentic search and memory retrieval graph so now they autonomously learn new skills and capabilities.
A major step toward a fully autonomous global supply chain where all the repetitive, error prone work is carried out by super intelligence AI agent swarms.

Probably the best write-up I've seen on Flock-related crime stats, the balance of privacy, and historical analysis.
Read the article here: https://www.richardhanania.com/p/flock-mass-surveillance-and-the-collapse
Richard Hanania: Did Flock cameras collapse the crime rate?
There are now about 120,000 Flock cameras across the US.
In the last few years, we have reached record low violent crime rates, and automobile theft is down 50%.
This probably isn't a coincidence. Surveys of police departments

Dev Writers Retreat + Unconference, Miami (2022)
One week of craft and community. A hurricane hit right in the middle of it but couldn’t kill the vibes.
What a crew. Great memories.

"A panic about a 'slave state' shouldn’t determine public policy. Intelligently regulated, ALPRs can be one among many tools in the toolbox of smarter—and therefore less severe—police forces. Dumping the cameras would just be foolish."
great piece @CharlesFLehman @TheAtlantic
Charles Fain Lehman: Flock Cameras are good. The online backlash is insanely dumb. It's possible to minimize the privacy risks while getting the very real safety gains that come from license plate readers and other cameras.
My latest in @TheAtlantic breaks it all down:
https://www.theatlantic.com/ideas/2026/08/flock-safety-cameras-ai-surevillance/688219/

It's exciting to see the launch of RVII and the focus on the YC ecosystem. I enjoyed this conversation with @richaberman and @spintopeyronel!
Shiv Verma: Early-stage investing has historically been difficult for everyday investors to access and even harder to understand.
This roundtable with leaders from Robinhood Ventures and Y Combinator offers a closer look at how Y Combinator works and how Robinhood Ventures Fund II is

What does a company that runs itself actually look like?
Our CEO @amasad joined @platformer to explain the self-driving company: an internal bot that acts as the brain of the business, engineers who ship far more code with agents, and a future where we use fewer apps while agents use them on our behalf.
Read more here:
https://www.platformer.news/replit-amjad-massad-interview-coding-design-jobs/?ref=platformer-newsletter

something that i've noticed in the ~4 months i've lived in sf is there's a small cohort of founders / engineers / ai researchers who all behave as if they're living 3-5 years in the future. this manifests in a lot of bruce wayne-type behavior. i've seen workflows of engineers who lie supine on the couch as they wisprflow commands to their agent swarms without ever touching a keyboard. i've seen people build out entire second brains that they orchestrate with openclaw / hermes agents. i know guys who spend $1m +/year on tokens.
so i think this conversation with alejandro maza, kavak's chief product and ai officer, with @astrange and @GEVS94 is super interesting because it 1) shows us what an entire organization built around living in the future looks like 2) kavak isn't even based in sf (or the us) so it's very cool to see how an organization outside of the bay area approaches ai from first principles. some takeaways:
- alejandro's governing question is what kavak would look like in 2035 with much more capable, cheap intelligence. kavak moved from functional specialists and transactions toward a persistent *agent per customer*, with a long-term goal of maximizing that customer’s lifetime value.
-every major model release should trigger a fresh model–harness experiment. kavak had developed a multi-agent framework, but a new model release rendered the company's legacy orchestration outdated, so they discarded 2 years of working infra and rebuilt everything around a simpler agent harness. the lesson is to keep asking: what's the minimum scaffolding the newest model needs to express its intelligence safely and at scale?
-kavak built an ai ceo for one city, and says it increased profits by 1.5x in its first month. customer satisfaction, inventory rotation, financing penetration, and other KPIs also improved. maza attributes the result to fields medal-level intelligence applied relentlessly to every number and customer: forecasting performance, assigning daily work, and collecting progress reports.
-in light of above points, alejandro's deeper thesis is that the organization, not merely the model or individual worker, should self-improve as new intelligence becomes available.
a16z: AI agents at Kavak sell the cars, underwrite the loans, coach the mechanics, and in one Mexican city, run the entire operation.
The Latin American used-car marketplace bet on agents three years ago. Today ~95% of interactions and transactions run end-to-end on AI: NPS tripled,

Watch Flow Studio’s Nikola Todorovic walk through one of many workflows possible in 3D Editor + Canvas, including how Marble by @theworldlabs can support the world-building stage.

Meta returns to open weights: Muse Glimmer, its first open-weights release since Llama 4, scores 35 on the Artificial Analysis Intelligence Index. It is a 30B-parameter model, and the first from Meta to be released under Apache 2.0
Muse Glimmer (high) arrives 16 months after Llama 4, scoring 21 points above Llama 4 Maverick (14), Meta's last open weights release. It sits alongside Kimi K2.5 (Reasoning, 36) and just behind Qwen3.6 27B (Reasoning, 38) and Ling 3.0 Flash (38), and creates a two-tier Meta lineup together with the proprietary flagship Muse Spark 1.2 (xhigh, 57)
@AIatMeta shared access with us ahead of public release for benchmarking. Congratulations to @AIatMeta, @finkd, and @alexandr_wang on the release!
Key Takeaways:
➤ Meta's open-weights line is back, under its most permissive license yet. Every prior Meta open release shipped under a Llama License; Muse Glimmer uses Apache 2.0, placing almost no restrictions on commercial use or derivatives
➤ Strong intelligence for its parameter count. At 30B parameters, Muse Glimmer scores 5 points above Gemma 4 31B (Reasoning, 30) at the same size, and effectively matches 1T total parameter Kimi K2.5 (Reasoning, 36) with 33x fewer parameters. Qwen3.6 27B (Reasoning, 38) remains ahead on the Intelligence vs Parameters frontier at a slightly smaller size
➤ Small enough to self-host on a single GPU, even at full context. Muse Glimmer is a 30B dense model (including a ~1.8B vision encoder) with weights at ~60 GB in BF16 and ~18 GB in 4-bit. It features a hybrid-attention mechanism with three sliding-window layers for every global layer, which holds KV cache memory use to ~1.8 GB (minimum) at its pre-extension 128K context. This means the model can run at full context on a single H100 at BF16 precision, or on a higher-spec MacBook or RTX 5090 at 4-bit, with more breathing room if the vision encoder is not required
➤ Agentic knowledge work is its weakness relative to its size class. Muse Glimmer scores 953 Elo on GDPval-AA v2, below the 1,000 human baseline and behind other models at its intelligence level, including the similarly sized Qwen3.6 27B (Reasoning, 1141), and Gemini 3.5 Flash-Lite (1141). Knowledge calibration follows the same pattern: its AA-Omniscience Index of -33 is low for its intelligence level, driven by an 82% hallucination rate (Qwen3.6 27B: 49%) rather than accuracy, where it matches its peers. Agentic tool use is the exception, with Muse Glimmer scoring 24% on Tau3-Banking, ahead of Gemini 3.5 Flash-Lite (18%) and Qwen3.6 27B (17%), among the best in its class
Other model details:
➤ Parameters: 30B Dense
➤ Context window: 128K tokens, plus extension
➤ License: Apache 2.0
➤ Openness Index: 44, indicating relatively high openness/transparency vs. other major open-weights models
➤ Pricing and output speed: at the time of release, Meta is not serving the model on their API; pricing and serving speed are TBC depending on third-party providers

The Democrats made a huge mistake letting themselves become the party of pronouns and land acknowledgements. They should just ditch that stuff entirely. Especially since it doesn't actually do any good.

This Meta campaign is a case study in strategic reframing, with 4 major examples:
1) REFRAMES THE AI RACE FROM “who builds it the best” TO “who distributes it to the most people”:
Instead of fighting over who has the most capable models, Meta is framing the race as being won by distribution. With open models, free products, consumer dna, and 3.5 bn users, Meta is best positioned here
2) REFRAMES COMPETITORS FROM “holding the technical lead” TO “holding too much concentrated power”:
Classic populist stance, with the insight that hostility to AI is largely hostility to being subjugated by an overlord class of AI perma-elites. The frontier labs’ biggest advantage (lead in capabilities) becomes a liability, while Meta’s huge user base is used to reposition Meta too powerful in its own right to being a check on the power of the frontier labs (the perfect foil)
3) REFRAMES SAFETY DEBATE FROM “more vs less safe” TO “optimists vs doomers”:
OpenAI and Anthropic have spent years focusing on AI risks, and many people see this as noble and responsible. By equating safety focus to doomerism, Meta repositions itself from being less safe to being more optimistic about the future, while making competitors look like tiresome hypocrites
4) REFRAMING META’S PORTFOLIO FROM “unfocused and capricious” TO “a future for everyone”:
In jumping between crypto, metaverse, AI, hardware, Meta has been accused of lack of focus and trying to do everything. But now they can position these motley initiatives as being unified by the coherent theme that Meta technology is “for everyone.” It’s not ADHD, it’s a mission to put frontier technologies into the hands of ordinary people
In the process, they’re also tapping into many parts of the zeitgeist: distrust of concentrated power, support for open models, permanent underclass anxiety, doomer fatigue
But Meta’s key strategic move here is to change the criteria for the AI race. Under the old criteria (capabilities of frontier models) Meta was seen as behind. Under these new criteria (widespread universal distribution of technology), Meta might actually have the strongest position in the industry
Mark Zuckerberg: I believe everyone should have access to superintelligence, and I wrote a long piece about Meta's philosophy and values for building a positive future for everyone. http://meta.com/thefutureisforeveryone

lot of talk about insecure sandboxes these days.
An elite security team just made a benchmark comparing them
Nebula Security: After the Hugging Face incident, have you wondered how easy it is to escape a modern agent sandbox?
We benchmarked eight open-source agent sandboxes to show how vulnerable they can be, and why frontier models can break out so easily. http://x.com/i/article/2086876544790933504