Remember: Fragile breaks under volatility Robust resists volatility Antifragile benefits from volatility

The Roman army rule: action removes fear.

Definitely is, i can see this on the ground.

Valued work per watt is the score

Loving the new Sol much more than the Astra 🙄 good job @OpenAIDevs

How?! Faster than a database query!

Today, I switched back to GPT Sol :)

Convex builders: has anyone used @convex itself as an MCP server, with @clerk auth, no Express/Node middleware in between? Would love to hear how you approached it, any gotchas, and whether you’d do it again. @jamwt @jamesacowling @Rasmic @waynesutton

VLMs are amazing. Astra made people realize it, but we still have a long way to go to make them useful in production at scale

Harnesses often get dismissed as just scaffolding, just prompt engineering, and not real research. But that couldn't be farther from the truth. The same model weights that score 30% on ARC-AGI score 95% with a better harness. So we gathered a group of researchers and founders working at the frontier to do a deep dive into the state of harnesses. We cover how we got to this point, the case for making your harness as expressive as possible, and what YC learned building an agent for every employee in the company. 00:00 - @FrancoisChauba1: Why harnesses matter 04:27 - Building an auto-researcher by accident 07:13 - A five minute history of harnesses 13:56 - Self-improving harnesses 18:35 - @sethkarten: Prime Agent, a self-improving RLM harness 21:50 - Context as an L1, L2, L3 cache 24:51 - From Turing machine to von Neumann computer 28:33 - Messaging between agents 30:04 - ARC-AGI results 33:09 - Emulator Bench and GPU kernels 37:30 - @JonSaadFalcon: OpenJarvis, personal AI on personal devices 38:26 - How far behind are local models 39:21 - The five primitives of a personal AI stack 42:47 - Letting cloud models optimize your local stack 43:53 - 800x cheaper than the cloud 45:58 - @josh__france and @jbellregan: QM, YC's agent harness for work 47:29 - A history of YC's internal agents 49:24 - OpenClaw and a fleet of 50 agents 51:04 - Pulling the brain out of the sandbox 54:43 - Letting the agent choose its own sandbox and model 57:16 - The grind tool: budgets on goals 58:50 - Agents don't understand social context

A gem from @ZachBDell on @davidsenra’s podcast: Two notebooks. Analog only. - Red: working in the business. - Black: working on the business. Friction by daylight. Architecture at midnight.

Spent all night debugging my transcription pipeline, convinced my code was broken. It splits audio into small segments to reduce latency. I moved from Gemini 3.6 to 3.8 because 3.5 Transcribe is weak on multilingual transcription (and double the price). Result: 12 segments. 3 worked first try, 2 on retry, and 7 had to fall back to 3.6 because 3.8 rejected benign audio as content_blocked. Same audio. 3.6 handled it perfectly. @GoogleAI please fix this. Safety filters shouldn’t make a transcription API this unpredictable. @OfficialLoganK @joshwoodward @osanseviero @DynamicWebPaige @patloeber @koraykv @demishassabis @shaneguML @sedielem @barret_zoph @GoogleDeepMind

Grok Bot internal best practices from 6 teams at SpaceXAI:

Might be a bit late, but I've just started using @bot , and it's fucking amazing!

It’s starting to feel a bit weird that 90% of my computer time is spent on a single app. And I’m still not over the rebrand, bring back the "Codex" name, ChatGPT just doesn't hit the same.

Gemini 3.5 Transcribe should be a major upgrade. After three days of testing, it feels like a step backward. I’ve used Gemini for transcription since Gemini 2.0 Flash and tested nearly every alternative. Gemini has often been exceptional at multilingual audio and speaker identification, which is why I was excited about this release. But 3.5 Transcribe consistently struggles with: • Switching between languages • Speaker diarization and timestamps • Foreign and RTL languages • Completing longer recordings, sometimes returning only 5 minutes from a 20-minute clip The surprising part? Gemini 3.7 Flash is currently more capable for transcription, diarization, and general audio understanding, and it’s 3 times cheaper!! @GoogleAI, what happened here? Is this the intended experience, or are fixes coming soon? I’d also love to hear from others testing it: are you seeing the same issues?

Agent builders: which framework would you choose today to ship capable agents inside a real application, and why? Production experience only, please. Share what you built and the biggest tradeoff you encountered. No benchmarks or vibes 👇 #AIAgents #AgentEngineering

A goal is a dream with a deadline.

In my previous company, we reached ~$2M ARR within six months on just ~$100K of pre-seed capital, paid in installments. We couldn’t raise from VCs. Years later, I angel invested in two startups that each raised $1M+ early. One is now at ~$50K ARR. When we invested, they had a pilot that looked like it could convert into a ~$3M ARR contract. It didn’t. The other is still at $0 revenue. Both have solid products. Lesson learned: 1. The best founders don’t just materialize leverage. They compound it. Capital should become talent. Talent should become product. Product should become customers. Customers should become more capital. 2. Execution is the conversion rate between opportunity and outcome. A great pilot, a great product, or a big raise means very little if you can’t convert it into revenue. 3. Greenfield still matters. It’s much easier to grow with a market than to spend years trying to force an existing one to change.