yutori.com/api
- 3 rounds of prompting (description, feedback+screenshot, feedback)
- GPT 5.4 xhigh still lacking contrast.
- Opus 4.7 max even with explicit description of content doesn't seem to get that the navigator agent has eyes. Output #1 was just a ball with no details.
I gave Claude Code & Codex a video of @yutori_ai Navigator logo spinning and asked for code to regenerate it.
Opus 4.7 max (left) vs GPT 5.4 xhigh (right)
GPT 5.4 clearly better.
Ground-truth / OG video in ๐งต
Two updates from Yutori:
1. We benchmarked GPT 5.4 on browser-use tasks
โข Matches/slightly-outperforms Opus 4.6 (+0.3%)
โข Big jump over previous OpenAI CUAs
2. Latest version of n1
โข Outperforms GPT 5.4 and Opus 4.6 (+3%)
โข 2.5x faster, 4-5x cheaper.
Most recent checkpoint of n1 vs Opus 4.6!
On Navi-Bench and Westworld browser automation benchmarks:
- Same accuracy
- n1 is 2.5x faster
- n1 is 5.6x cheaper
Try it out via the Yutori API.
Fun chat with Evan O'Donnell about the similarities between training robots and web agents, managing context for agents that run for months and years, the future of the AI-first web, and ideal form factor for embodied AI.
www.thetimes.blog/p/agents-ne...
Maybe coding is just amortized inference for LLMs.
Maybe the reason we write programs down to files is just to save inference costs.
Why?
Because the web is a mess. And websites are fundamentally built for human consumption.
If you want to automate everything that a human can do with a browser, then you have to perceive like a human.
Stay tuned for more in the coming weeks!
yutori.com/blog/the-bit...
The bitter lesson for web agents
The last 1 year has taught us a new bitter lesson that we think others are not yet grokking.
Agents that *look at the web like humans* (screenshots of sites) navigate and generalize better than agents that read code (HTML, DOM).
Solved: robustness to paraphrasing and false premises, OCR, world-knowledge based reasoning.
Open: spatial reasoning, data-efficiency, learning compatible representations.
As part of the award ceremony, VQA team presented a recap of vision-and-language research over the last decade โ solved problems, progress, and open-challenges for mutimodal LLMs.
Lots to be done. Thank you to all our collaborators and the research community for this recognition!
Fun-fact: the T-shirt I'm wearing is an inside joke about the quality of 2015 models.
However, every few years we rediscover the lesson that on difficult tasks, VLMs silently regress to being nearly blind.
x.com/DhruvBatra_/...
VQA challenge series won the Mark Everingham prize at #ICCV2025 for stimulating a new strand of vision-and-language research.
It's extra special because ICCV25 marks the 10-year anniversary of the VQA paper.
When we started, the idea of answering any question about any image seemed outlandish.
Anything by Ted Chiang
I dunno man, Dagger is cool.
The problem with โAI slopโ isnโt the AI โ itโs the slop.
People act like AI is the issue, when itโs actually part of the fix.
If we're honest: most of what we make, most of the time, is slop by our own standards.
Thatโs the generatorโdiscriminator gap in creative work that Ira Glass talks about.
Somebody is a fan of Abundance
It is so refreshing to see conferences innovate on the reviewing model and run actual experiments (!) as opposed to fighting change.
Good. Autonomous interface locomotion is the fundamental robotics problem of our time. The more the merrier.
My entire robotics career has led to this.
The answer to many "why X?" questions:
Because the laws of physics do not prohibit X and the forces of biology gave us curiosity.
The web is the ultimate boss-level for agents โ dynamic, non-deterministic, and noisy; some mistakes are inevitable and so far, every agent fails eventually.
Yutori is building superhuman agents for this ultimate digital environment.
Join our waitlist for early access to our product!
yutori.com
๐๐ฆ๐๐ ๐ข๐ง๐ ๐ ๐ฐ๐จ๐ซ๐ฅ๐ ๐ฐ๐ก๐๐ซ๐ ๐ง๐จ ๐ก๐ฎ๐ฆ๐๐ง ๐ก๐๐ฌ ๐ญ๐จ ๐๐ข๐ซ๐๐๐ญ๐ฅ๐ฒ ๐ข๐ง๐ญ๐๐ซ๐๐๐ญ ๐ฐ๐ข๐ญ๐ก ๐ญ๐ก๐ ๐ฐ๐๐ ๐๐ ๐๐ข๐ง.
Where teams of AI assistants coordinate to book flights, manage budgets, or file paperworkโproactively surfacing insights and correcting errors.
Only problem โ no one knows how to build AI agents that actually work.
I started something new last year with a wonderful group of people. We showed a demo in Jan.
Today, weโre telling our story โ show before you talk!
๐๐ฆ ๐ข๐ณ๐ฆ ๐ณ๐ฆ-๐ช๐ฎ๐ข๐จ๐ช๐ฏ๐ช๐ฏ๐จ ๐ฉ๐ฐ๐ธ ๐ฑ๐ฆ๐ฐ๐ฑ๐ญ๐ฆ ๐ช๐ฏ๐ต๐ฆ๐ณ๐ข๐ค๐ต ๐ธ๐ช๐ต๐ฉ ๐ต๐ฉ๐ฆ ๐ธ๐ฆ๐ฃ โ one of humanityโs greatest inventions and a a mess overdue for an overhaul.
yutori.com
Ah, understood. No idea about the tracing of that meme.
Seems like the ultimate thing to rally around, no? To the extent there is any purpose, what's the alternative?
I'm already there for low-stakes queries.
Where's the skepticism coming from? Now that web search and citations are in there, isn't it easy to verify and thus become more confident?
