Thomas Friedel

Head of Machine Learning @plantixapp

[Thomas Friedel @thomascygn]

@doodlestein https://t.co/f6NypsSQjP

[Thomas Friedel @thomascygn]

@doodlestein my wokaround was to just use a low framerate like 0.1 (basically audio only with a frame every 10 sec). However, they had a bug that caused the resulting transcripts to be truncated. May have been fixed with new models. Haven't checked in a few months.

[Thomas Friedel @thomascygn]

@Trumpyla @iScienceLuvr I trained a classifier on LLM output and human text and it can distinguish the two really well. I then built a browser extension that runs that model on your machine and filters out the slop on twitter: https://t.co/pXnakiA0NT

[Thomas Friedel @thomascygn]

@MLStreetTalk @VictorTaelin can't you close the feedback loop for example by having gemini *watch* the video and report any of those micro errors?

[Thomas Friedel @thomascygn]

@perplexity_ai @PPLXDevs can I use this for custom document collections ? And what is the latency?

[schmitthub @schmitthub]

you should always install rootless docker whenever possible, it can be a pain the ass but worth it. Don’t ever let an AI agent run unattended on a system that doesn’t have rootless docker setup. It’s a classic privesc vector and in my experience most LLMs will try it fairly quickly when hyper fixating on completing the task.

https://t.co/m1vtEVM073

[Ibragim @ibragim_bad]

first update:

Claude Opus 4.8 – xhigh on march-may 110 tasks: 56.4%

gpt-5-xhigh: 62.7% – $2.25
gpt-5.5-medium: 58.9% – $0.98
Opus 4.8 - xhigh: 56.4% – $2.02
Opus 4.7 – high: 53.1% – $1.32
Opus 4.6 - high: 47.8% – $1.29

more open-weight models to come in ~1-2 weeks

[Thomas Friedel @thomascygn]

@giffmana @hassonofer_ S3 has unfortunately high egress fees ($82 / TB). Not great if you actually need to download your backup. Would pick something else: https://t.co/8QwKiBMLjR

[Vik Paruchuri @VikParuchuri]

Announcing Surya OCR 2:

- 650M params
- 83.3% olmocr bench score (top under 3B)
- 87% on internal 91-lang benchmark
- 5 pages/s on RTX 5090
- Runs on CPU, GPU, MPS

[Thomas Friedel @thomascygn]

@doodlestein you can also use gemini which can understand videos natively

[hardmaru @hardmaru]

For over a decade, we’ve accepted that end-to-end backprop is the only way to train deep networks. But holding the entire network in memory all at once is why AI training is hitting a resource wall.

We found a new way to break the network into blocks and train them independently. The trick? Treating the network’s forward pass like a diffusion model denoising a signal.

This reinterpretation slashes the memory needed to train deep models. In our #ICLR2026 paper (https://t.co/PK5h0mqQSo), we matched end-to-end performance across ViTs, DiTs, and LLMs. We did this while training just one isolated block at a time.

[sunil pai @threepointone]

the window on ai-assisted writing being acceptable has closed (unless models change drastically). I have a couple of https://t.co/nj3DDFEnZw posts that I'm going to leave there. t'was good because I wouldn't have posted them at all. but I don't think I'll do it again for a while.

[Thomas Friedel @thomascygn]

@madaomoshiroi source: https://t.co/GMeJbRwqMj

[まだ面白い @madaomoshiroi]

火山噴火の瞬間を上から捉えたドローン映像、これは凄すぎる

[Thomas Friedel @thomascygn]

@omarsar0 Built something similar with Gemini. We got video recordings of presentations and needed to ingest them to a RAG system. So it created a PDF with screen shots + textual descriptions + audio transcript below the slide.

[Boris Cherny @bcherny]

Big fan of teaching more people the basics of using Claude Code in an accessible way.

So much of the world has not yet used agents. There's a lot of opportunity to level the playing field and expand access.

[Thomas Friedel @thomascygn]

@Suhail I made a browser extension that hides AI generated ones:
https://t.co/yEABJeBGWa

[Thomas Friedel @thomascygn]

@omarsar0 how does it deal with compaction? does it do this at sensible points or does it just fill the context window?

[Thomas Friedel @thomascygn]

@danluu Curious what kind of project are you running with such high token spend? Some major porting effort? A fully autonomous software factory with dozens of agents in parallel?