#Digital #Native & #Innovation Enthusiast. #Design Thinker. Software Developer. Keynote Speaker. CTO @ MEETYOO
Berlin, Deutschland
Joined December 2009
- Tweets594
- Following609
- Followers148
- Likes2.4K
I bet Ox Alpha is a Gemini model.
1M context
Video input
Smells like gemini.
Especially video input kinda unique to gemini models isn‘t it?
Tim Adelmann retweeted
We benchmarked Mistral OCR against other frontier and open-weight models on ParseBench 📊
For a model at its price point, it is quite competitive!
- It wins on semantic formatting - understanding strikethroughs, superscripts/subscripts, title hierarchy, links
- It is competitive on content faithfulness (reading order + hallucinations + omissions) and visual grounding (bounding boxes)
- It does ok on tables and doesn't really have chart capabilities.
Of course, some of the frontier models + OCR providers like Azure Doc Intelligence + AWS Textract are a bit more expensive.
Check out our full leaderboard on ParseBench: parsebench.ai/
Can an llm in a agentic application decide by itself to summarize the context to safe context?
Can an agent do context engineering on its own?
Tim Adelmann retweeted
Apple should buy Anthropic at any price.
"Claude Code just SAVED my marriage!"
I gave Claude Code 512 GB of HEAVILY corrupted wedding footage from a close friend of mine, and within 10 minutes, recovered it all.
Two days ago, I got some bad news from a buddy telling me how the only copy of their wedding videos was corrupted to do a severe fault in the memory card.
Just like that, someone's happiest day turns into a nightmare. The footage was practically gone, nothing was able to read the memory card, and the only option was to pay thousands of $$$ or let it go.
Until I had an idea...
What if I plugged in the memory card and let Claude Code have a go at it.
It first made a copy of the corrupted files onto my computer using PhotoRec as to keep the original drive untouched. For context, I have never done anything like this before, it did this itself.
It looked at memory slices from the IDR, then began using FFMPEG to reconstruct the SPS, IDR, P/B, etc. all without me touching it once.
I come back to the full 512GB of uncorrupted video I can send back to my good friend. The power of Claude Code is insane.
I feel like MCP needs some metadata for proper context engineering.
Eg. Context masking (removing tool outputs after some turns to free context) would require some information which tool can be/should be masked.
How should this be done in the current spec?
My year in code. @cursor_ai
There are a couple of days left in 2025. I‘m confident i will reach the 1T until the end of 2025!
The most amazing thing this year was the rate of improvements of models and IDE!
Thank you
Hot take: The only good reason to use an SLM is if you're running on-device.
For cloud/enterprise, the future isn't "Small Models." It's Smart Routing. The true achievement of GPT-5
Don't optimize by shrinking your intelligence. Optimize by routing your traffic.
Two massive deals: OpenAI's $1B Disney IP grab for Sora + Netflix's Warner Bros bid.
Why? AI video abundance is here. Attention is scarce—IP is the ultimate moat for beloved slop (Star Wars floods incoming).
Netflix fights TikTok/IG too, not just Prime.
I always thought putting 2 of the same icons next to each other is bad UX.
Well I‘m definetly confused. Always press the wrong (X) … liquid (gl)ass
AI agents need data, but you can't let them run wild in your production database.
Our solution: A secure, in-flight sandbox. The agent can now explore, analyze, and visualize customer data autonomously and safely.
Css anchor is around the corner.
Finally one of the most anticipated APIs in the web.
I guess @jaffathecake is to he thanked for the FF support.
The biggest myth in AI right now? That one "super model" has all the answers.
We tested 17 LLMs on a video analysis task. They didn't agree. Instead, they formed "opinion clusters," each finding different highlights.
The lesson: The smartest AI isn't a single genius, it's a committee of experts. This has huge implications for building AI products.
read the analysis:
contenthub.meetyoo.com/resou…
Unpopular opinion: Multi-agent "swarms" are often just technical debt in disguise.
We tried the Planner > Executor > Replanner architecture. It was a hot mess of hallucinations and loops.
We switched back to a single, high-capability model. It outperformed the swarm immediately.
"Sorry" for the lack of posts. I’ve been busy with this. 👇
Current status at @meetyoo:
• Full dedication to the code (yes, even as CTO).
• 10+ hidden features in the works.
• Zero "vaporware".
We don't build for the hype. We build for the customer.
You ever forgot in which branch you buried a change?
I do all the time.
No more.
I can now create a list of diffs for a single file in all my recent branches.
`bash branchSearch.sh path/to/my/file.ts > changes.txt`
ALT Branch Search with diff
Tim Adelmann retweeted
Europe is so back.
No more cookie banners, if member states approve of the proposal.
theverge.com/news/823750/eur…
