Founder of @hedra_labs, making visual intelligence fast Former PhD @Stanford SVL working on visual reinforcment learning, simulation, and embodied AI
San Francisco
- Tweets 514
- Following 668
- Followers 2,249
- Likes 309
We're excited about an upcoming launch. The cost of GPUs are becoming astronomical. @AMD has been a huge unlock for us, as we've managed to get amazing performance across multimodal models (VLA, WAM, Image, Video).
We're working directly with some of the best companies in the world, like @krea_ai, to optimize and post-train models for improved performance.
Feel free to reach out if you're looking for compute -- we have a ton of capacity coming online and will help you optimize your model.
Krea 3 is here, and 6x faster on Hedra.
Built for reference-driven image generation and editing:
→ Combine character, outfit & scene references
→ Explore angles & compositions
→ Restage products in new settings
→ Edit details in your images
See it in action ↓
1
1
11
500
ChatGPT is great for distribution -- excited to bring more of our infrastructure offering into ChatGPT so you can deploy compute, open source models, and more from codex!
Hedra is now in ChatGPT.
Add it from the plugins directory, attach a product photo, and get a commercial without leaving the chat.
Every shot in this ad was generated with Hedra in ChatGPT 🫧
2
2
518
How I'm maximizing my various token subs:
* Codex + Astra -- daily driver for eng
* Claude + Fable 5.1 -- code review (usage limits too low for anything else)
* Hermes/Pi/Grokbot + Grok 4.6 via supergrok for admin
Protip, you get supergrok for *free* for all affiliates of X business accounts.
Altogether about 225 USD/month, not bad!
3
1
8
924
Why are Github runner services (@useblacksmith, @depotdev) so unreliable?
I have a PR that we revert every few days which switches us between them when jobs start hanging.
1
8
391
Best side project to contribute to right now! Zig is so fun to review.
Things I worked on while on the plane:
Kitty image support
Steering/queuing support
Attach mode over ssh
fx v0.0.6
• Better Auto mode (cheaper & faster)
• Approval workflows up to 23% faster
• Smaller macOS binary: 6.12mb (-0.26%)
• Remote HTTP MCP support
• More reliable Codex, Grok & MCP sign-ins
• Many bug fixes
We've also documented MCP integrations with @neondatabase, @supabase, @prisma, @usekernel & @browserbase.
Thanks to @sachinadlakhaa, ackingyou & @mjlbach
Full: fx.sh/changelog#v0.0.6
1
17
2,774
The main thing that frustrates me about remote sandbox development is the additional terminal latency. I still type a lot (amid Wispr'ing), and it's obviously noticeable even on servers near SF.
I'm working on extending fx (from @vercel_dev) with support for attaching to a remote session over tailscale RPC. It feels *really* good.
I've always been sensitive to typing latency. During COVID in my neovim maintainer days, I spent an entire week profiling every bit of the LSP client and lifting the performance sensitive components (mostly diff tracking and json parsing) into C.
1
8
1,034
Zig has one *major advantage* over Rust in the AI coding agent era, it's extremely fast to compile.
I was having 4-5 fx agent iterate overnight extending it with < 1s compilation time which is huge for testing.
Already have a working prototype of a plugin system via WASM that I'm super bullish on, I *may* abandon Pi.
Introducing fx, a tiny, open, native coding agent from Vercel Labs.
Originally an internal tool, fx is a harness and CLI written in Zig, optimized for research and embedding in larger systems. Today, we're open sourcing it.
fx is built on three principles:
1. Fast. A single native binary, no runtime to install. It cold starts in 10µs and does no unnecessary work or I/O before accepting input. fx is the answer to "how fast can a coding agent be?"
2. Light. The 6.3MiB binary uses single-digit megabytes of memory at baseline, made for instant installation and embedding in resource-constrained environments and agent sandboxes.
3. Open. Apache-2.0, model and provider agnostic, suitable for local and cloud inference. Its small core extends through skills, plugins, and MCP.
Minimalism is an obsession throughout the entire harness: system prompt, tools, features, binary. The goal was to keep context usage and time to first token low, and make fx optimal for model benchmarking, sandboxing, evals, and gyms.
You can use fx directly or embed it as infrastructure. The CLI feels more like a Unix shell than an IDE in the terminal: it preserves scroll history, produces minimal output, and uses complex TUI rendering very, very sparingly. Programmatically, 𝚏𝚡 𝚊𝚜𝚔 --𝚓𝚜𝚘𝚗 gives structured output, 𝚏𝚡 𝚊𝚌𝚙 connects to editors and other clients, and WebAssembly can even run the whole thing inside the browser (see: fx.sh/try).
Privacy is a design constraint: no product telemetry, sessions and usage stay local, and no source code or prompts are shared with any endpoint other than inference. With local inference and auto-updates off, fx is fully hermetic.
fx is experimental. Use at your own risk and expect frequent changes. Chat with us on X (fx.sh/xchat) or file issues (fx.sh/issues).
𝚌𝚞𝚛𝚕 -𝚏𝚜𝚂𝙻 𝚏𝚡.𝚜𝚑/𝚜𝚎𝚝𝚞𝚙.𝚜𝚑 | 𝚋𝚊𝚜𝚑
fx.sh/
7
16
749
.@signulll says we're still in the stone age of AI tools:
"I don't think most people are utilizing them anything beyond the basics."
"We're in the stone ages of how people view and perceive and use these things, even though there's a billion people utilizing them, but they're not utilizing them to the full capabilities."
"The number one challenge... is how do we make this stuff, the power of the models more easily accessible and useful in terms of what they can do."
"I think this is happening with agents. It's happening today, but it still seems very primitive and very inaccessible to a lot of individuals... we need to make this stuff much more easily accessible and useful for individuals."
53
28
417
186,478
The new gpt image 2 destroys anything that is not nano banana and legitimately has a sense of humor. Joint VLM/generation models are insanely powerful.
2
1
7
701
Today we're launching Omnia - Hedra's first general purpose world model.
A lot of video models today are still feeling like uncanny valley. The movements are exaggerated, the expressions make you a bit uncomfortable, and you can tell something's wrong even if you can't articulate what.
We spent a lot of time post-training Omnia specifically to fix all of this. And I'm happy to report that when we ran testing, the results were noticeably different from what you can get with other models right now.
Try it now at hedra.com
7
6
58
6,323
Michael Lingelbach retweeted
MVPs are easy. Building something people want to use is hard.
@mjlbach, CEO + Founder of @hedra_labs, did just that, reimagining video as a collaboration between humans and AI.
In his Start Here feature, he shares how he built the MVP and what’s next: pleaseplatforms.substack.com…
4
25
4,124
Michael Lingelbach retweeted
Finally, I'm getting somewhere with my AI avatar, powered by Hedra Studio and its new Character 3 model.
And with Kling 1.6, Veo 2, Minmax, Flux, ElevenLabs, and more AI models all bundled into one interface, there's so much more Hedra can do for you.
39
8
131
6,936
The @Cloudflare eng team is so cracked. We reported a bug in the streams API and got a 1 day turnaround on the fix going out to prod.
gg @eastdakota 🫡
4
13
78
8,050
Michael Lingelbach retweeted
Presenting: The Bridge. An AI Short Film utilizing Veo-2.
I’m really proud of this one, as my goal (as always) is to push storytelling, performance, and narrative in this emerging art form.
Every shot here utilized Veo-2, although there were a few post-generation tricks here and there. Hedra was used for the lipsync.
Writing, Sound, and Editing were done by me.
Yeah, I’m pretty proud of this one.
Per usual, I’ll have a full run through of the production on the channel in a few days!
80
91
687
48,344
