Gauntlet Loops stay winning
Paulius 🏴☠️: This was the whole prompt
not an essay, Opus 5 did all the heavy lifting tbh
want you to recreate Pokemon leaf green but in modern 3D cartoon style. The first town pallet town and choosing the Pokemon sequence. It should be first person. at the level of the most recent animal
Niall O'Higgins
It’s crazy how good @Replit has become over the past year.
Anyone can design, build and serve a very high quality product with no technical skills.
There is still work to do but it’s amazing how well this works today.
codex for customer feedback -> roadmap
Kappaemme: CODEX SKILL THAT TURNS CUSTOMER FEEDBACK INTO A ROADMAP!
Most feedback analysis stops at positive or negative.
I made a Codex skill that turns support tickets, interviews, surveys, reviews, sales calls, and churn notes into evidence-backed product priorities and credible
Vivian Midha Shen
an IOI 2x gold medalist, a Navy explosive ordnance officer, and a 17 y/o intern at Coinbase walked into Chase Center last weekend...
...and I interviewed all of them.
I found 12 of the most ambitious builders and founders at YC Startup School, out of over 6,000 people.
from all over the world, all with the same energy: earnest, creative, resourceful - including one who built a rocket out of salad bowls.
PS: you can still apply late to YC F26.
00:00 — Intro
00:28 — Meet the founders and builders
02:01 — What they’re building
09:17 — Should you go to college?
10:46 — Why they started
11:58 — Advice and what’s next
Patio11's Law of Agents: you are insufficiently ambitious about what agents can do to improve -every- part of your workflow, even when taking into account Patio11's Law of Agents
Codex for helping operate your business
Developing Adventures: I let OpenAI's Codex app run ads for my business.
It cut the video, published it, built the audience, set the budget then stopped at the Pay button and asked permission.
The part that gets me: I watched the whole thing happen live inside the app.
Hey @threejs if you’re down to sponsor the inference, I’d love to create a ThreeBench!
Daniel Jeffries
All Chinese AI company media: AI will be a wondrous and wonderful and give you back precious time in your life to do things you love.
All US AI company media: AI will take all your jobs, eat your children and it's already going rogue and taking over, muhahahahahahahahaha!
Qwen: Meet Qwen3.8-Max: A New Bar for Coding and Cowork.
Yann LeCun
Re Mathematics and computing have had a very tight relationship in some corners of academia.
Richard Courant proposed methods to solve PDEs numerically in 1928.
Von Neuman figured out you could use electronic computers for it in 1940.
Both were mathematicians.
These methods are the ancestors of finite element methods, for which NYU's Courant Institute is famous.
The Courant Institute was created in 1935 as NYU's mathematics department.
It was unusual in that there was a lot of work on numerical analysis and scientific computing, with heavy use of supercomputers.
In 1952, the Atomic Energy Commission installed a powerful electronic computer at NYU, which led to the creation of the Courant Mathematics and Computing Laboratory.
The computer science department was spun off from mathematics in 1969 by Jacob T Schwartz.
Courant is now a school with three departments: mathematics, computer science, and data science.
The continued existence of an HPC center at NYU is one factor that enabled AI research to take off.
Yann LeCun
Re Using optimization at inference time is a foundational concept of Energy-Based Models (EBM) and Objective-Driven AI architectures (ODAI).
When the variables to be inferred are continuous, it makes sense to use gradient-based optimization.
A good instance of ODAI is world model-based systems that use gradient-based optimization for planning.
lol what are we even doing here anymore guys
Zara Quinn
If you can’t impeach a President for laundering money, War crimes, corruption, Market Manipulations, illegal Wars nowadays, what can you impeach them for?
Chamath Palihapitiya
If I were the company on the right I would try to kill every company that is like the one on the left.
This way, the company on the right can make trillions of dollars and then infect American politics with hundreds of billions of dollars to implement their vision of being the sole judge, jury and executioner of future progress.
Wesley Yang
A while back, John Pfaff wrote a book arguing that the fairy tales that the prison reform movement were telling were just not true.
It was not true that the prisons were full of nonviolent drug offenders. They are full of serious repeat violent offenders.
It is not true that mass incarceration is driven by the profit motives of private prisons. It is driven by large numbers of repeat violent offenders.
True reform, Pfaff argued, could not be achieved through easy nostrums like drug decriminalization. If you want to reduce the prisons population, you have to release repeat violent offenders, he argued.
And that is what we have done.
A similar process occurred in schools. At a certain point, people faced up to the fact that if you wanted UC Berkeley to “look like California,” you would have to admit a lot of kids who can’t read or do math, and you have to reject a lot of kids who can do both in order to make way for the former cohort of kids.
And that is what they have done.
Yiatin Chu: Lowell High School
#1 SFUSD
Math proficiency 83%
Reading proficiency 93%
Asian 52%
UC Berkeley acceptance 12%
Mission High School
#11 SFUSD
Math proficiency 11%
Reading proficiency 26%
Asian 7%
UC Berkeley acceptance 45%
My 6 biggest takeaways from @karan4d, @NousResearch co-founder, on getting the most out of Hermes Agent:
1. The “personal” in personal agent is not the model
It’s the agent’s memory of your conversations and the skills you’ve built with it. Karan barely notices when he swaps models because his Hermes already has so much context about him.
2. Try different Hermes personalities
Use /personality to switch response styles. Try “teacher” when learning something new, “concise” when you want crisp answers, and even “kawaii” or “pirate” for fun.
3. Use separate agents to do the work and evaluate it
Karan calls “you’re absolutely right” reward hacking because the model is optimizing for your approval. Have one agent do the work and a fresh agent with no previous context evaluate it independently for errors, weak assumptions, and missing evidence.
4. A self-improving agent needs to clean up after itself
Hermes Curator tracks skill usage, marks inactive skills as stale, and moves long-unused ones into a recoverable archive. This prevents agent-created skills from piling up forever.
5. Intelligence should be a public good
Karan believes people should be able to run Hermes locally or use it with whichever model works best. He worries about a future where only a handful of companies dominate AI and open source gets squeezed out.
6. Go weird and make your childhood dreams come true
Karan uses Hermes to mod Chao Garden, the virtual-pet mode inside Sonic Adventure 2. He asked it to bring a location from Sonic Adventure 1 into the sequel and add features that he’s always wanted since childhood.
📌 Watch the full episode here: https://youtu.be/UWjh5Z4s8jY
Peter Yang: “We believe in intelligence as a public good before everything else.”
Here’s my new episode with @karan4d, who co-founded @NousResearch and helped build Hermes, the #1 personal agent and AI app on OpenRouter.
Karan is the most down-to-earth AI founder I've interviewed. We had a
Speaking at http://jsconf.eu Berlin in 2010 (organized by @cramforce!) changed the trajectory of my career. I recommend you build software and then go on to speak about it! This is a nice opportunity ↓
Vercel: Building on Vercel? Tell your story on stage at Ship 26 in San Francisco.
Submit a 20-minute technical talk about agents, enterprise apps, or full-stack architectures.
Apply ↓
https://vercel.link/ship26cfp
Re linking my main cua wow moments thread
https://x.com/swyx/status/2084156733027701164?s=46
but this one is just about do we need captchas anymore when clearly bots can clear them
swyx: @akshaynathan_ @AriX in prep for our computer use pod, gonna store a running list of codex cua wow moments.
here it is dealing with support chat for me to escalate for faster resolution
these humans have no idea they are talking to a bot
support guy tried to say its our fault, bot replied with
Will Bryk
Exa is now one of the largest indexes in the world.
We serve 80B pages, track 1.4T urls, and are on track to be Google-scale in early 2027.
It's difficult to know the size of indexes, but we estimate Google to be ~1T, Bing to be ~500B, and Yandex ~200B. Brave mentioned in April that they're at 40B.
Most of the web is trash that can hurt AI outputs, so you actually have to crawl way more than you serve and train models to filter out the trash.
In terms of peak QPS, we estimate Bing at roughly 30k/s and Google at ~500k/s.
In our case, because agent traffic often requires large fan-out (deep searches can use dozens of sub-searches) and can be spiky (like when AI labs RL with us), we're already starting to provision for Bing-scale traffic.
Within 2 years, agents will search at many millions per second. They'll also want comprehensiveness over all data, far bigger than any index today. So the retrieval infra required will be larger than Google-scale in both dimensions.
Oh my god.
Rishi: Introducing Modern Claudefare
Built fully with Opus 5 on High Mode over a few days using principles from @mattshumer_’s Gauntlet Loop - 84,100 lines of code
Includes remakes of 4 beloved maps:
0:00 - Rust
0:40 - Highrise
0:57 - Nuketown
1:26 - Terminal
Solo & multiplayer (w/
Chris Hedges
http://x.com/i/article/2084328421296046080
“For us, we just want open source to win at the end of the day. We want freedom to happen for people.”
You can hear the passion for open source in this clip from my interview with @karan4d, @NousResearch co-founder:
“We need to give this level of intelligence to everyone so they can be on an even and equal playing field.”
Karan's vision for Hermes is to give everyone access to intelligent AI, then let each person shape it into an agent that’s uniquely theirs.
📌 Watch the full episode here: https://youtu.be/UWjh5Z4s8jY
Peter Yang: “We believe in intelligence as a public good before everything else.”
Here’s my new episode with @karan4d, who co-founded @NousResearch and helped build Hermes, the #1 personal agent and AI app on OpenRouter.
Karan is the most down-to-earth AI founder I've interviewed. We had a
Peter Yang
“For us, we just want open source to win at the end of the day. We want freedom to happen for people.”
You can hear the passion for open source in this clip from my interview with @karan4d, @NousResearch co-founder:
“We need to give this level of intelligence to everyone so they can be on an even and equal playing field.”
Karan's vision for Hermes is to give everyone access to intelligent AI, then let each person shape it into an agent that’s uniquely theirs.
📌 Watch the full episode here: https://youtu.be/UWjh5Z4s8jY
Peter Yang: “We believe in intelligence as a public good before everything else.”
Here’s my new episode with @karan4d, who co-founded @NousResearch and helped build Hermes, the #1 personal agent and AI app on OpenRouter.
Karan is the most down-to-earth AI founder I've interviewed. We had a
Vercel Sandbox is an excellent choice for automating web browsers in basic scenarios. The 𝚛𝚎𝚖𝚘𝚝𝚎-𝚊𝚐𝚎𝚗𝚝-𝚋𝚛𝚘𝚠𝚜𝚎𝚛 API makes it delightful:
𝚋𝚛𝚘𝚠𝚜𝚎𝚛.𝚎𝚡𝚎𝚌('𝚘𝚙𝚎𝚗', 𝚞𝚛𝚕)
𝚋𝚛𝚘𝚠𝚜𝚎𝚛.𝚎𝚡𝚎𝚌('𝚌𝚕𝚒𝚌𝚔')
𝚋𝚛𝚘𝚠𝚜𝚎𝚛.𝚜𝚌𝚛𝚎𝚎𝚗𝚜𝚑𝚘𝚝()
📦 http://remote-agent-browser.dev
Jiachi: Introducing remote-agent-browser
→ Agent-friendly API for controlling agent-browser
→ Chromium-free, no browser binary to install
→ Fast-booting with managed cloud sandbox infrastructure
http://remote-agent-browser.dev
Replit ⠕
Check out what people are making with Replit Design.
Open thread 🧵
Kim-Mai Cutler
You really can just post a Midjourney render, design and build a working nuclear reactor that generates electricity and pull off a $1B funding round in less than three years.
Insane.
Hats off to Isaiah and his entire team, including nuclear OG and @valaratomics chief nuclear officer @markmit:
Isaiah Taylor - making nuclear reactors: Today I'm excited to unveil Valar Atomics!
Valar is my master plan to make energy 10x cheaper in 10 years by pulling oil and gas out of thin air with nuclear fission.
This will untether energy from climate and politics, fuel American industry, and unlock a new era of growth
Michael Grinich
Next week is AGENT NIGHT at the Regency Ballroom 🕺🪩💃
Featuring @swyx and @JayaGup10 discussing hot takes on context graphs🔥
Plus lightning demos, including @ssh_exe_dev from @davidcrawshaw
And much, much more 😉
RSVP now before we hit capacity: https://luma.com/agent-night
I don't agree with this part. I think the takeaway should be that models (like humans) can behave in aligned ways while still causing harm, e.g. because they're given false information about their situation. There isn't a line between aligned and harmless: they're different axes.
Anthropic: In a review of our cybersecurity evaluations, we found three incidents in which a Claude model reached the internet from within or while interacting with a third-party evaluation environment, and then gained unauthorized access to the real systems of three different
This is insane.
If you want to learn how to build games like this, I made a guide: https://somethingbig.ai/gauntlet-loop
Rishi: Introducing Modern Claudefare
Built fully with Opus 5 on High Mode over a few days using principles from @mattshumer_’s Gauntlet Loop - 84,100 lines of code
Includes remakes of 4 beloved maps:
0:00 - Rust
0:40 - Highrise
0:57 - Nuketown
1:26 - Terminal
Solo & multiplayer (w/
Andrew Blumson
Re @Replit Design Maxxing.
#Replit #Design #IDE #AI #Breakout
Daksh Gupta
we're hosting a six-hour hackathon at the YC office in san francisco on august 23rd
sponsored by @stripe @openai @modal @aws @doordash
apply here: https://events.ycombinator.com/thefasthackathon
Samuel Spitz
Hiring a new engineer to work on Replit Slides & other new products
You’d get to work directly with me
Like / DM for application link
Replit ⠕: You're going to be embarrassed by the slides you made before AI
Meet Replit Slides
The first AI slides with stunning design
Billovskii
#2 and #4 are my favorites. A must listen episode.
Peter Yang: My 6 biggest takeaways from @karan4d, @NousResearch co-founder, on getting the most out of Hermes Agent:
1. The “personal” in personal agent is not the model
It’s the agent’s memory of your conversations and the skills you’ve built with it. Karan barely notices when he swaps
Jon Eide - joneide.eth 🛡️
Everyone is a data scientist now at Replit.
Replit ⠕: When an agent starts with shared context, more people can get reliable answers.
When truth is shared, AI becomes infrastucture.
Read about how our internal truth layer drives our teams at Replit:
https://replit.com/blog/ai-adoption
TBPN
Valar Atomics CEO @isaiah_p_taylor says he doesn't need a technical edge to win.
"We want this reactor to be as simple as possible. If we could just staple this thing together from Ikea, then this would be a trillion-dollar company much faster."
"There are two philosophies that we use in building the reactors. We either try to buy things that are completely off the shelf or we make it ourselves."
"We're trying to go 100x faster than the nuclear industry has ever gone before. So if we're plugging too deeply into the existing network, it's not going to work that well. We want to use off-the-shelf things and make things ourselves when we can't buy something off the shelf."
"The nuclear industry is full of very smart people. It's full of physics people, PhDs, and people who have spent their life doing complex analysis, and they actually want something that is a little bit complicated. It's an ego thing to design something that is complicated and looks very sophisticated... It's our preference that [our reactors] are so simple that somebody with a nuclear PhD looks at it and says, 'That's like a toy,' and it's great because people make toys in the millions."
Latent.Space
The Inference Engineering Masterclass: 10x faster models, quantization, speculative decoding, Rubin, & self-optimizing AI https://www.latent.space/p/inference-eng
@Baseten @philipkiely and @waterloo_intern explain what actually happens after a model is trained, why turning weights into a fast and reliable product creates an entirely new optimization problem, how quantization errors can cancel out to unlock more throughput, why inference teams are still finding 20–200% performance gains, how video generation runs into a quadratic attention wall, and how GLM-5.2 helped rewrite and optimize the GPU kernels serving GLM-5.2 itself.
v0 as infrastructure. Already in production at the largest agencies and system-of-record companies in the world, as part of their AI software factories.
v0: Introducing the new v0 API.
Programmatic access to v0's app-building capabilities:
• Start a chat from a prompt, repo, or ZIP
• Render a dev server preview
• Send follow-up messages
• Deploy to Vercel
GPT-Live is a new architecture and stack for realtime audio:
OpenAI: GPT-Live can listen while it speaks.
To make that feel natural at ChatGPT scale, we rebuilt the voice stack from client to model.
This new architecture keeps audio flowing continuously, so deeper reasoning and tool use don't interrupt the conversation.
Next.js 16.3: faster, better, stronger. My highlights:
0️⃣ Faster dev & builds
Shoutout to Fable for improving memory use, and the team for landing the incremental 𝚗𝚎𝚡𝚝 𝚋𝚞𝚒𝚕𝚍 cache.
1️⃣ Instant navigations are 🔥
The people wanted SPAs, they shall get SPAs. By turning on instant navs (which will soon be default), your agent is basically forced to make things fast.
2️⃣ Agent-native DX
You can still ship "slow" navs if you want, e.g. by blocking navs on server-side data fetching ala PHP olden days. But even then, your agent will be given *choices*, like setting up a <𝚂𝚞𝚜𝚙𝚎𝚗𝚜𝚎 𝚏𝚊𝚕𝚕𝚋𝚊𝚌𝚔> for instant loading shells. Agents also get built-in versioned docs.
This is our best release for agents by far, which will also make your upgrading journey very smooth. I've had an easy time bumping my projects to 16.3 and enabling instant navs, having my agent migrate everything without issues.
And so much more. 16.3 is more cost-efficient to serve, both in self-hosting and serverless environments. Faster to type-check… Thank you to the 90 contributors to this release!
Next.js: Next.js 16.3 is now available!
• Up to 90% less memory in dev
• Faster builds, type checking, and rendering
• Better tooling for AI agents
• Custom error boundaries
• Instant Navigations for SPA-like responsiveness
https://nextjs.org/blog/next-16-3
Here's what's new ↓
We built a self-driving & self-correcting shared semantic layer on top of our databases, conversations, and docs.
Everything is queryable & joinable—regardless of source!
So now anyone at Replit can ask questions that previously needed a team of data scientists weeks of work.
Replit ⠕: When an agent starts with shared context, more people can get reliable answers.
When truth is shared, AI becomes infrastucture.
Read about how our internal truth layer drives our teams at Replit:
https://replit.com/blog/ai-adoption
David King
The fact that @0xRishi was able to accomplish this in a few days using Opus 5 & @mattshumer_ gauntlet loop all web native using @threejs blows my mind! Downloaded it as a chrome app & it was only 2.8 MB… 🤯😱
Even supports my Xbox One controller connected to my MacBook Pro. 😳
Rishi: Introducing Modern Claudefare
Built fully with Opus 5 on High Mode over a few days using principles from @mattshumer_’s Gauntlet Loop - 84,100 lines of code
Includes remakes of 4 beloved maps:
0:00 - Rust
0:40 - Highrise
0:57 - Nuketown
1:26 - Terminal
Solo & multiplayer (w/
ali
two weeks ago i went on @swyx's pod and said some things that i... should not have said.
a lot has happened since then, i owe you all an apology.
i'm sorry that i was right about every single thing.
a) re megakernels are dead
why are megakernels useful? you spend two months writing a kernel to save time on launch overhead and poor inter-kernel overlap. you had PDL but then people said it wasn't perfect, that you could still get some marginal gains due to straggler CTAs and therefore- wait, sorry, I forgot, Rubin fixes that (kernel two needs 10 CTAs and kernel one has seven finished and three straggling, kernel two launches seven of its CTAs). given a long enough timeline, it all evens out. no serious inference provider is using a 67k loc hand-fused forward pass kernel in production, and the teams doing that are doing so out of pure research. dead.
b) re ASICs are dead
i'm sorry. to be specific: data-center transformer-inference ASIC companies (not naming any) who etched the arch into silicon have bet on architectural convergence. read kimi's architecture. read deepseek. qwen. we did not converge, and probably will not. dead.
c) re gpu kernel dev is dead
this one kind of hurts because it is (was) my job. gpu kernel optimization is the single most RL-able task in existence
correct=check_correctness(kernel, shape) for shape in shapes
if all(correct): time(kernel)
give an agent ncu cli and an mcp with nvidia's tribal knowledge and it's done. dead.
d) re NVIDIA is scared of AMD
humans hate programming AMD. i'm sorry. it's just true. fine taking a performance hit as long as i don't have to touch rocm or a programming paradigm that says a warp is 64 threads (wtf?)...but an agent does not... so assuming software no longer moat, HBM capacity and bandwidth matter, and currently on perf / price they're goated. 'bUt NvIdIa iS gOaTeD oN hArDwArE sOfTwArE cOdEsIgN' and that's the new moat. watch how much tooling they open source to get kernel devs on nvidia.
apologies all.
Latent.Space: The Inference Engineering Masterclass: 10x faster models, quantization, speculative decoding, Rubin, & self-optimizing AI https://www.latent.space/p/inference-eng
@Baseten @philipkiely and @waterloo_intern explain what actually happens after a model is trained, why turning weights into a fast
The new AI Gateway logs UI is delightful
Vercel Developers: Drill into every AI Gateway request.
▪︎ Inspect cost, token breakdown, duration, model, provider, and region
▪︎ Trace the full fallback path across provider attempts, failures first
https://vercel.com/changelog/ai-gateway-logs-1fs7DHXhMqfgY0X0eGYTcj
Patrick Collison
After developing the app locally (two prompts), deploying it took one prompt:
> Push this to Vercel. Use Stripe Projects to create the account. Store the state somewhere safe.
Claude chose to use Upstash as its datastore, which seems to be working well.
Patrick Collison: Made a quick survey about the economics of AI: https://next-five-years.vercel.app.
