Daniel Ou

Polishing up Ghostnotch, you can now launch agents directly from a directory + liquid glass. Releasing on Github soon🤓

@creativelyange huggingface robot 👀 just bought all the electronics for it huggingface.co/docs/lerobot/e…

@jun_song less engagement bait posts too

New X algorithm got significantly better. Now I don’t see any unrelated videos like World-cup or shitty viral videos from my timeline. Now it’s full of Local AI or tech related contents. Amazing

@theo well I never hear people complaining about ChatGPT limits so.....

Do we get resets for ChatGPT going down or only Codex?

anyone try connecting Bambu Lab 3d printers to agents for automated print farms? 👀

@creativelyange anything is worth it for more compute 🙏

driving an hour just to secure a mac min rn

@xmangonic i think this is good advice regardless of age

If you're in your 20s rn Start using AI every day

@mweinbach least obvious GPT-5.5 design

Netflix home screen is a card on a background... I would recognize a GPT-5 design anywhere

@sauravroyX yep subagents are cool, id be really careful with superpowers though, better to use a native implementation also be careful with spawning too many subagents or dumb subagents (haiku)

Recently, I find subagent driven development (superpowers has one built in) is pretty amazing Let top tier models spawn subagents of varying levels of weak models depending on the tasks I let Fable spawn a bunch of Haikus, some sonnets, and Fable reviews everything with its large context window

I think the days of using a strong model to plan and a weaker model to execute is over. We should also be using capable execution models as they are doing more complex things nowadays such as orchestrating subagents or computer use.

@LuminaXspace grok is honestly quite a decent model, maybe overshadowed by the 5.6 release quite excited to see what SpaceXAI + Cursor brings in the future

🚨 Grok 4.5 tops the new Long-Horizon-Terminal-Bench Researchers tested frontier AI models on 46 extremely long, real world terminal tasks: • Grok 4.5 ranked #1 overall with a 19.6% pass rate • Grok fully solved 13 of the 46 tasks • Each task consumed approximately 9.9 million tokens on average • Agents completed roughly 231 episodes per task • Average execution time reached 85.3 minutes • Tasks covered software engineering, scientific computing, multimodal analysis etc SpaceXAI cooked with Grok 4.5.

@0xSero this the future! not even a bad performance loss at 1 bit... we need more local models that can run on <16GB as thats what most consumers have

8gb VRAM best model for your hardware 1 bit gguf doing better than models 25x its size!?!?! This will even run in 4gb (: huggingface.co/prism-ml/Bonsa…

@theo hiccups = resets 🙏

“It is possible there are some hiccups soon” Oh no

@siddsubh0 yeah 5.6 sol for both its expensive but its great for both planning and long horizon execution

@imnotchalk True, do you have any plans for doing that?

@AgenticLazyness @Adidotdev that is a valid point, google has the most potential financial wise if they can release the more competitive models as well it would put them in the best position over the other labs in the long run

@imnotchalk @Adidotdev One could argue anthropic is making profit, but their fiscal year hasn't yet ended, and it would be their first year making profit, IF that happens. So google is making profits, its the most used model in general, why would they need more

gemini 3.5 pro drops july 17 > 2 million token context window > deep think reasoning > $1.25 input / $10 output per million tokens that's cheaper than sonnet 5 and a bigger context than fable 5 3 days from now the entire AI leaderboard reshuffles anthropic is about to feel real pressure

@AgenticLazyness @Adidotdev very benchmaxxed, have you heard of anyone actually using Gemini models professionally? I haven't. They may be making good money, but their models are truely not competitive with the other labs at this point, maybe 3.5 pro will change that.

@imnotchalk @Adidotdev people forget that each release from every company gets #1 , including past google models. people in this platform just have brain damage and forget what happened earlier this year. I believe their slow releases are due to sustainability. Google is the single company making $

@zanlib0 @figma yeah, elytra its super nice, quality is good, lightweight and portable battery also lasts forever

@imnotchalk @figma this is elytra? how do you like it?

If you use a trackpad who hurt you

@saneord anthropic deadass buggin these days 😭 blaming openai for ts

wtf an ai company have anything to do with graves is anthropic saying ai kill us and they are our saviour?

@kimmonismus how will it stack up to fable? they have to make it competitive with sol which is an akward spot as it also needs to be weaker than fable

Claude Opus 5 on vertex means release is kinda imminent. My prediction: They are removing Fable 5 from the subscription plan this coming Sunday, but will release Opus 5 shortly beforehand to appease everyone.

@LuminaXspace bullish on Deepseek 🙏

🚨 DeepSeek is preparing a major expansion of its AI model programme Deepseek is reportedly seeking another enormous funding round: • New discussions could value DeepSeek at around $71 billion • Funding would support a dedicated DeepSeek data centre • The company intends to acquire significantly more AI chips • Future development is reportedly focused on autonomous AI agents • Agent systems require far more inference compute than ordinary chat models DeepSeek has not announced a V5 model, but the infrastructure expansion suggests its next generation could be far more agent-focused. Are you looking forward to Deepseek V5?