Sinan Koparan

I research AI, machine learning, and LLMs by building with them. Agents, RAG systems, experiments that probably shouldn't work but sometimes do. Follow along for papers worth reading, tricks that survive contact with real data, and the stuff hype cycles miss.

MModern software engineers are now coding from their phones. Just run claude --dangerously-skip-permissions and /remote-control.

@sethkarten Seth, this looks so cool. Without checking out the paper just yet. I’m just curious how the agent was able to provide low latency outputs?

We've solved another piece of the generalist gaming agent puzzle! Mario requires different skills than Pokemon: reactive navigation, spatial reasoning, and safe exploration rather than long-term memory and zero-sum reasoning. Check out our new paper on finetuning VLMs to beat Mario via PPO. 👇

@bindureddy Hopefully! Definitely some new capabilities to come as well. You have to remember that Gemini is more multimodal than other model providers

Gemini is over due. Expect a really good Flash model Something that beats GPT 5.5/Opus 4.7 medium but is considerably cheaper and faster

@fabianstelzer Hahahaha

when Claude Opus 6 tells you to "stop spiraling and go to bed" 😵‍💫

@kimmonismus How come they say what model or algorithm was used? They just say AI.

1/ Holy: Astronomers just pointed an AI at NASA data from 2.2 million stars. It found over 100 hidden planets, including worlds so extreme they shouldn't even exist according to current theory. I love it. Lets break it down and explain what it means 🧵:

You do not need to write code to experiment with models. Several providers offer free browser-based playgrounds that give you direct access to their models with full control over settings. Google AI Studio lets you interact with Gemini models directly in your browser.

Hugging Face Spaces is genuinely one of the best things for AI engineering projects right now. You can go from a model or prototype to a live, shareable app in minutes. No heavy infra setup, just build, deploy, iterate. It’s become the default playground for shipping and testing AI ideas quickly.

@gabrielchua Which country are you in now?

@Taniyatweets_ Give me the krabby patty formula, thanks!

1M impressions complete ✅ Need 4M more to unlock monetization No breaks till then

Armenia is investing $4 billion on AI. They’re building a mega hub powered by thousands of NVIDIA Blackwell GPUs.

@Its_Nova1012 Who’s going to be the ones building AI?

AI is already writing better code than many devs today, so what are your future plans?

@BhartiAnsh2007 Just push it whenever you feel like a significant feature has been completed. It could even be a config file.

Genuine question how often should i push my codes to GitHub ? Like i push 1-2 time after completing coding at last !

No hits on llms.txt after 5 months of it being available on my site.

@Jashanx_gill @sama I noticed a change after the Molotov incident. Not a nice thing to have done to you. Also, his blog addresses a lot of things that many people were concerned for years.

Is @sama becoming everyone’s friend? My feed is starting to feel more human.

@sama @LinghuaJ @OpenAI Don’t sleep on the marketing. It was really well too!

@LinghuaJ @OpenAI team made an amazing product and model, i think its really just about that

Amazed by how fast the trends change over night - Codex @OpenAI all over the feed. You gotta win the community. @sama

@kimmonismus Super keen to see new model capabilities. Hopefully it isn’t a letdown!

This SVG surfaced in an AI Studio A/B test window, and it’s hard to believe it’s actually vector-generated. likely from Google’s new Flash/Pro model. So freaking hyped for Google i/o! 2 weeks to go!

@hnshah Can you give examples of the products being built? I'd be curious to know

I'm obsessed with running local LLMs. Been working with an engineer to build product(s) that are 100% local. A new model that came out recently instantly improved the quality of our product. We live in really interesting times.

Well, it was their limited resources initially that forced better ideas with Deepseek R1. They ended up focusing on efficiency, cleaner data, and smarter training instead of just scaling everything up. But yes, with more compute, they probably would have had better models faster.

If DeepSeek had the compute & memory resources of Anthropic, we’d probably already be talking about DeepSeek V7… maybe even V8.

Most social media teams are bottlenecked by production. Livestreamers solved this with the clipper model: dedicated clippers turn hours of live content into a constant stream of short-form posts across TikTok, Reels, and Shorts. Video understanding makes this replicable at scale. Models can ingest long-form video, detect high-engagement moments, and output timestamped clips with captions and hashtags.

Slopsquatting is a supply-chain attack where AI hallucinated package names become malware traps. Flow: AI invents a dependency (e.g. asyncgraph-streamer) Attacker publishes it on PyPI/npm Developers install it, trusting the AI suggestion AI hallucination → real package registry → malware