Eric Litman

Speechless.

@4gb7xvbwqx You know we’re all waiting for your TestFlight versions…

@ericlitman @ericlitman I totally agree!

One of the best games of all time

@BarryBarrywu @nwcarlson ^^ this please. The Mac Sonos app is trash.

@nwcarlson Native AirPlay 2 output is on my list to tackle; wired + Bluetooth speakers sync great in the meantime.

macOS still only plays audio through one output at a time. 🙃 So I built Tutti: tick any combo of speakers, AirPods, and displays in your menu bar → same sound everywhere, in sync. With real volume + mute per device. #buildinpublic #macOS #indiedev

@4gb7xvbwqx You, sir, are a hero among men.

I used Codex to port Command & Conquer: Red Alert to Mac and Android. No emulation. The original engine now compiles natively to ARM64, with macOS and Android platform layers. Codex wrote the code, built the binaries, ran them, played the game, found bugs, and fixed them. All in a /goal loop. (Inspiration from @ammaar ) Full open source below.

@sundeep It’s made its way into my daily nighttime ritual.

The best way to fall asleep is to /goal your agent on a difficult task!

@philippberner Not ready to say happy yet, but started using this today: blacksmith.sh

Anyone happy with their CI setup? if yes, what is it?

A few patterns we frequently use with Fable 5: Use Fable 5 as an "advisor." An executor (Sonnet 5) calls Fable 5 for guidance. Most tokens are billed at the lower executor rate.

@rabi_guha @_jshmllr iOS rendering would be 👌

This is still rendering React. Didn't have bandwidth to write the native renderers. When we saw gemma on cerebras out eyes lit up. We have been anticipating this day when we can do subsecond rendering reliably and this is the first time it felt real On a side note: fable should be able to one shot the renderers - just need someone's bandwidth to read the code

Imagine you never needed an app again. Every screen on your phone, generated live the moment you ask. No app store. No downloads. No predefined screens. You ask for dinner, and the screen to order it gets written - right then, just for you. We're calling it AppLess.

@biogerontology Awesome, Alex. Good luck with this one, huge if/when you pull it off!

Hello, my dear friends, we have some breaking news today! Our rentosertib (novel target that has never been in clinical trials before we picked it using AI, novel molecule designed using true generative AI with the published broadly used AI platform, targeting fibrosis and biology of aging) entered Phase III trial in Idiopathic Pulmonary Fibrosis. This is a major milestone in the life of every drug hunter - the ultimate test. To my knowledge, in our AI drug discovery industry, no other company took the drug for a novel target with a novel molecule into Phase III before so it is a big milestone for the entire industry. Of course, at Insilico we are no longer dependent on this program - we have 13 INDs and many clinical trials are ongoing internally and run by partners. But this program is very special as we also demonstrated the senomorphic properties of this drug and built the entire pipeline of molecules around this target. Please Wish Me Luck! Luck is the most important superpower in everything we do and if we succeed with this program and also demonstrate the antiaging effect - all of us will benefit. Godspeed, my friends and thanks to everyone who helped us along the way!

@thsottiaux No mobile notifications from the desktop app

What is something that you feel is surprising that Codex still can't do well and we should have gotten right a while ago?

@chamath Or they’re trying to force their team to dogfood Grok internally

Tesla is one of the smartest, cracked and most advanced engineering companies in the world. If they actually did this, then it is likely verifiably true that a dollar above $200/week is waste.

@trevin Fable prompting has evolved from Opus days. Get Fable to rewrite those for Fable! and given the cost of Fable output tokens, encourage it to be concise 🙏

With Fable back, one thing I’m doing a lot is ce-brainstorm and ce-plan using Fable, then switching models to Opus for /lfg or /ce-work. working great.

Can’t express just how big of a deal this is

@whatdotcd 👌

Friends who use personal agents (Hermes / Openclaw / codex / cowork etc) daily reply to this so I can send you something cool to try

Now you can develop an oral fixation atop your agent addiction

🚨COMING SOON: SUPERSONIC FLIGHTS ✈️ The @FAANews is clearing the way for supersonic technology to take the skies at MACH 1 SPEEDS ‼️ The airspace is about to get faster… Welcome to the Golden Age of Travel 👇

Introducing Claude Science, a new app designed with every stage of research in mind. Artifacts traced to their code, environments managed on demand, and 60+ optional scientific databases that you can connect. Available now in beta.

Between this, the work @colossal, and others are doing, there’s a potential future where the entire human reproductive cycle requires no humans at all

@no_stp_on_snek ming sharing config / runtime parameters?

verdict up front: Qwopus3.6-35B-A3B-Coder-MTP looks like a pass in the practical agent lane, not because it beats Ornith everywhere. it doesn't. the interesting part is where it wins. Qwopus is stronger on the boring agent stuff that makes a model feel usable: legit-request compliance, integrity under pressure, multi-turn orchestration, large code deliverables, sustained debugging. that sounds less flashy than "reasoning," but it is the stuff that breaks real coding agents. does it do the allowed work or stall behind fake prerequisites. does it keep state across the loop. does it finish the artifact. does it keep moving through fix-test cycles without turning every step into a speech. the poison result is the important caveat. Ornith still wins context-poison resistance, 85 vs 70. so no, Qwopus is not the cleaner "misleading human" model. if your main fear is the user injecting a false premise mid-stream and the model quietly rewriting history around it, Ornith is still the stronger specialist there. but Qwopus wins the practical execution cluster. and for local coding agents, that cluster is not secondary. most turns are not grand reasoning moments. they are inspect file, edit code, run test, read error, continue. a model that handles those turns directly without over-gating is useful. my read: Ornith is still the more cautious long-reasoning specialist. Qwopus is the better do-the-work local agent. the trade is real. Qwopus gives up some poison robustness and broad engineering judgment. but it buys cleaner execution behavior, lower reasoning drag, and better completion on the stuff that turns into actual code. not magic. not a strict superset. a strong practical coding worker. tested/scorecard on the official card, benchmarks courtesy of yours truly.

@KyleHessling1 Great work. Getting ~80tok/s with ~1s TTFT on an M2 Mac Studio Ultra, with higher quality results than I've seen on any other Qwen derivative (including Ornate). Any chance of a MLX version? Feeling greedy, would love to see if this can get up to 100tok/s.

Good morning y'all! Qwopus-3.6-35B-A3B-MTP-Coder is live! All GGUF's will be populating over the next few hours! It's a lightning-fast MOE with the coder curriculum recipe. Similar to the 27B coder, it shines with thinking disabled, offering significantly faster wall time for similar, and in some cases superior results to same-sized thinking alternatives! With thinking disabled, it goes toe-to-toe with the new Ornith 35B MoE across a huge eval suite (performed by @no_stp_on_snek), edging it on the coding trajectories and decisively on speed and cost, even though Ornith was run with thinking enabled. See the model card for the full test results, and shoutout to Tom, @no_stp_on_snek, for thoroughly evaluating the model for us before launch! With MTP and thinking disabled, along with the MOE speed, it runs so quickly in harnesses like @opencode that it almost feels instant @ 253 tps on my 5090. No 8k tokens of thinking before a coherent output is actioned. This is especially useful in long contexts, where the base models will progressively start thinking for tens of thousands of tokens before replying. Compared to the base models with thinking off, the coder curriculum really advances the no-think frontier. Especially in terms of how creative it can be. Run temp hot as usual, 0.85-1, and make sure your harness isn't overriding the temp setting of your server at runtime. If you want to use it to its full ability, I would recommend giving it very thorough prompts. I have been using it in opencode, and I have been blown away by the results it generates autonomously with chunky prompts. Please see links to the demo's Aether Dominion (RTS Game), and a slide deck presentation the model made about itself that turned out beautifully, links in comments below! I am getting results on this incredibly fast local model (with thinking disabled) that I couldn't get in some thinking frontier models over a year ago. Open source is accelerating fast, and in light of recent events, there's never been a better time to get your local AI workflows tightened up. This MOE would be a great one to play with, and it's also a great one if you don't have much VRAM because it can run fast offloaded partially to system memory! All of that said, please give it a run with thinking off and build something you'd like to see. We'd love to see your results and any feedback on specific use cases in the comments below! Also, thanks so much for 5k followers, you all make up such an enjoyable and knowledgeable open source community, and I am so blessed to be able to collaborate and discuss this research with all of you. I can't express how grateful I am for every comment. As always, I will try to reply to them all! If we ever get monetized on X, I will put every penny into buying more hardware for our lab! Have a blessed day, my friends, looking forward to your thoughts! huggingface.co/Jackrong/Qwopu…