chu2bard

Washington kid, Dartmouth econ/CS, built a sportsbook in 47 states from a dorm room Rebet then Trivelta, now building @utopiadata

[chu2bard @chu2bard]

.@grok what's your take on the data bottleneck in physical AI, is it a real constraint or just a talking point?

And can a startup like @utopiadata, paying people to film specific places and turning that into verified 3D meshes for robotics labs, actually have product-market fit in that vertical, or is the market still too early?

[chu2bard @chu2bard]

@ravikiran_dev7 dm me or open your dm, have an inquiry

[chu2bard @chu2bard]

If you're building anything that has to move through the physical world, my DMs are open.

[Utopia @utopiadata]

Utopia has raised $1M.

Thanks to Gate Labs, Monarch Group, Ascentis Capital and RyzeLabs for backing us.

Everyone agrees physical AI is next but it has almost no real-world data to learn from.

We turn the phones people already carry into an economic network that collects it.

[chu2bard @chu2bard]

@jumperz Rate limits and refusals are a product decision, not a bug list. Anthropic optimized for enterprise trust, not power users venting on Twitter at 2am.

[chu2bard @chu2bard]

Machines don't care about your architecture either, they just execute instructions, the "human" part is whoever has to maintain it at 3am.

[shadcn @shadcn]

Engineering is for the machines. Design is human.

[chu2bard @chu2bard]

@GojoInWeb3 @utopiadata ✌️

[chu2bard @chu2bard]

in sf next week for @infodrivercap AI data event

Compute got cheap, models got good and now everyone's discovering the same thing: robots/models are starving for data nobody has collected yet

come say hi if you're around✌️

Join the event: https://t.co/fSrPJ38PzU

[chu2bard @chu2bard]

@jakezward 100% visibility across the board makes me suspicious the sample prompts were too generic, "form builder" isn't exactly a hard query to rank on.

[chu2bard @chu2bard]

Would love to know how much of the release cycle is actual new features versus quietly patching cache invalidation edge cases nobody talks about publicly.

[Tim @timneutkens]

We're the Next.js team. Ask us anything!

We recently shipped Next.js 16.3, and we’re excited to talk about what’s new, how we approached the release, where we are going, and what we’ve learned while building and maintaining Next.

Ask us anything about Next 16.3, App Router, React Server Components, performance, caching, upgrading your applications, contributing to the framework, or what it’s like to work on Next.js.

[chu2bard @chu2bard]

Running lots of agents in parallel doesn't fix bad output, it just multiplies it faster.

- 80 agents for 12 hours sounds impressive until you count what's mergeable
- Throughput without review capacity is just noise generation
- The bottleneck was never model count, it's someone reading the diffs

[chu2bard @chu2bard]

@rwalk_xyz CC0 is the right call, most "free" template sites bury a trademark clause that makes you attribute them forever. Does the MSA cover liability caps by default or leave that blank?

[chu2bard @chu2bard]

Free until you stack five of them with different caps and discover your cap table math doesn't add up until a priced round forces the reckoning.

[Ankit Gupta @agupta]

YC invented the SAFE a decade ago, simplifying the process of raising seed money for founders across the world.

Today, we’re releasing a SAFE tool to make it even easier to raise your seed round on SAFEs. And it’s free!

[chu2bard @chu2bard]

The color coded cable runs matter more than the GPU count. At that density one mislabeled bundle costs you a week of downtime, not a training checkpoint.

[Joe Hansen @joehansen]

Grok 5 is still deep in training on the massive Colossus II cluster.

Roughly 220k high-end NVIDIA GPUs are already online, with more being added.

Power draw is sitting around 1 GW and climbing.

The model is targeting the 6 to 10 trillion parameter range, which puts it in a completely different league from anything currently public.

Training has been running since at least January.

Elon recently confirmed it should arrive before the end of 2026 and will be trained on the full 25-year SpaceX engineering data set.

That combination of extreme scale plus proprietary real-world engineering data is the real story here.

This is not another incremental Grok 4.x update.

It is the much larger architectural and compute jump happening in the background while the smaller releases keep rolling out.

When it lands it should give SpaceX a serious internal edge on engineering, simulation and design problems that current models still struggle with.

Cursor is already running on the same cluster, so the jump in capability should show up there too.

No firm date yet, but the ambition on both compute and data is very real.

[chu2bard @chu2bard]

Anyone actually running zero data retention in production, does it hold up once you add tool calls and multi step agents, or does logging creep back in through the back door?

[chu2bard @chu2bard]

@vral 40% average across what mix though, that number means nothing without knowing how much was gpt-4o traffic that should've been on haiku anyway.

[chu2bard @chu2bard]

5 of 5 attempts to count as solved is a brutal metric, 60% at that bar means Gemini is actually consistent, not just lucky once.

[Philipp Schmid @_philschmid]

Gemini 3.7 Flash just took #1 on @ArtificialAnlys new AA-AnalystAgent.

AA-AnalystAgent evaluates against 80 real-world quantitative analysis tasks across 14 business and scientific domains (finance, healthcare, hydrology, government appropriations).

The Agent run inside an isolated Python 3.12 sandbox using AA's open-source Stirrup harness. They receive reference spreadsheets (.xlsx) and documents (.docx) alongside standard data libraries (pandas, polars, openpyxl, scipy, PyMuPDF) to inspect schemas, write scripts, handle edge cases, and calculate final figures.

Gemini 3.7 Flash:
• Accuracy: #1 with 60.0% pass^5 (70.5% pass@1, 77.5% pass@5)
• Speed: 1.32s per task (fastest)
• Cost: $0.54 avg per task (middle)

[chu2bard @chu2bard]

6.19GB at 1-bit runs, sure, but check the perplexity numbers before calling it usable, IQ1_S on a 27B model is a coherence lottery.

[Lonely @Lonely__MH]

🎉好消息！ 8G 内存 Mac 也能部署 Qwen 3.8-27B 了！

@UnslothAI 刚刚整了个大活，发布了基于 Qwen3.8-27B 的 Dynamic 3.0 超级量化版本！

以往 27B 这种体量的多模态大模型，至少需要 30G 以上的显存/内存才能勉强带得动

但这次，通过极致的 1-bit / 2-bit 量化（UD-IQ1_M / UD-IQ2_S），模型体积直接被压缩到了 6G~8GB 左右！

虽然是极限压缩，但它依然保留了多模态能力，还自带思维链深度思考！

运行方式👇🏻

1. 羊驼 Ollama 一键运行（适合小白）
---
ollama run https://t.co/AN1Vy6QhTt
---

2. llama.cpp 终端运行（推荐极客）
---
brew install llama.cpp llama cli -hf unsloth/Qwen3.8-27B-GGUF:UD-IQ2_S
---

[chu2bard @chu2bard]

Copy paste three.js templates versus npm packages, different tradeoffs entirely.

- A 150KB procedural file an agent can edit beats a 12KB import it can't touch
- Bundlers don't care about prompt friendly code, they care about tree shaking
- Once you paste it in, you own the bugs, no more waiting on a maintainer

[chu2bard @chu2bard]

@aidenybai Depends which console update broke, the minimap or the disconnects. Both were live today, pick your chaos.

[chu2bard @chu2bard]

Memory across self-hosted sandboxes is the part people underrate until the second run forgets what the first one figured out.

[ClaudeDevs @ClaudeDevs]

We shipped three updates to Claude Managed Agents.

First, we made it easier to use memory with Self-Hosted Sandboxes.

Any work done in a Self-Hosted Sandbox can be saved to memory.