Looking forward to it. I'm also a big fan of https://vincentwoo.com/2023/07/20/tunnel-vision-an-unauthori..., thanks for keeping the internet interesting.
2,813 karma · joined January 1, 2022
> This scan is made possible by recent advances in Gaussian Splatting. This is an emerging technology that lets us quickly create very detailed models just from photographs. For this model (or splat, as we call them), my friend Daylen and I flew our drones around Sutro Tower at a respectful distance for an afternoon until we had collected a few thousand photographs.
> I then aligned these pictures in free software called RealityCapture. Alignment is the process that teaches the computer that a bunch of points in different images all actually correspond with the same point in real life. Then I used another piece of free software called gsplat to produce the 3D model itself.
I'm assuming this was done similarly. Very cool.
> On Sept. 30, 2025, a Flock vice president logged into the Dunwoody PD’s Flock system and looked at a camera in the children’s gymnastics room of a private community center here in Dunwoody.
And there is a lot being implied in this sentence.
My experience has been quite the opposite. I was using Claude Code almost exclusively this winter/spring and swapped to Codex earlier this summer. It took no time whatsoever to switch. And before Claude Code, I was using Cursor. Same story.
[edit: Oh and there was also a brief interlude with Conductor, though I think they're more or less just serving the underlying Claude/Codex harness]
I asked Codex to generate an article with a high score and then asked Codex to (reverse?) hill climb that score. The original generation scored 97% and then the optimized one scored 1%. Both are pretty bad and read like slop.
https://gist.github.com/wbew/8a2bd6686bf875210f2244ac8ea65bf...
It's clearly a dupe of Claude/Codex products (Codex especially, styling-wise), but I think Kimi's goal here is simply to appear on feature-parity with bigger labs. I doubt they'd want to invest much into designing a UI for "the future of agentic work" from the ground up.
Their bread and butter and claim-to-fame will continue to be low-cost, near-frontier models, for which they need to at least appear to have some app layer to plug into.
And, in some ways, duping is a great way to show people that even the app layer can be commoditized.
I used to talk myself out of it all the time, but have recently just been going for it. It's been great.
Open base models that can be fine tuned on Tinker is a great business model IMO. You (i.e. an enterprise) can own your own model & have it perform frontier-or-better at your task at potentially much lower cost and Thinking Machines gets to be your essential infra/service provider in this world.
Also,
> Inkling-Small matches or exceeds its larger sibling on many benchmarks — the result of improvements we made to the pre-training data and recipe for the smaller model.
Very cool! Excited to see the next generations of Thinky models.
> The report was initially closed as Informative and out of scope. After we challenged that determination, HackerOne reopened the report, reproduced the issue, and confirmed that the details had been delivered to Cursor. And then everything stopped. Requests for updates went unanswered, additional follow-ups received no response, escalation through HackerOne produced no meaningful engagement, and direct outreach to Cursor leadership yielded the same result: no response.
Really unfortunate. I don't understand why there's such a lack of response on the Cursor side.
[replaced a Threads link for a better one]
Ok you have my vote.
> And two, this strawman jump from, oh hey, it’s a fancy autocomplete, smart compiler, better search engine, to it’s gonna like own the whole light cone bro like if you aren’t in SF and at the right parties there’s gonna be like a flash of light in the sky one day and you’re not even gonna know what happened but everything just Changed.
Haha, OP has a way with words.
In a way, both these emotional extremes (FOMO & the singularity) are just tools being used to continue driving the massive CapEx behind LLM improvement. Hate to love it? Love to hate it?
> Over a six month period, we adopted a city-wide switchback (also known as crossover) experimental design, alternating between this treatment and the control (unaltered) routing algorithm over consecutive days to appropriately measure the effect of this intervention
> Averaged across cities, we observe a median increase of around 2% in driving speeds on targeted segments, corresponding to a median decrease of 0.5% to 1.0% in fuel consumption rates
The cities were: Atlanta, Boston, Chicago, Los Angeles, Miami, New York, Philadelphia, Salt Lake City, San Francisco, Seattle.
The data and code is also available (https://github.com/google-research/google-research/blob/mast...) from the paper (https://www.nature.com/articles/s44284-026-00443-x).
Kudos Google! Nice to see this kind of work. That said, let's just build more trains?
This is almost certainly a reaction to GPT 5.6, which IMO is better positioned than Fable from a cost perspective. It's crazy to see how the tides change between labs every few months.
Easier said than done to be honest, especially if there are many people (and their agents) pushing code. It’s hard to keep up these days.
> We assessed how reliable current measures are for trying to find microplastics in blood. And what we found is that lipids and fats will give you a false positive for polyethylene.
> We worked with an architect, and we built the lab pretty much from scratch. [...] So we ended up going with stainless steel. It was the only way to not have any plastics.
> I don’t think we’ve got really good evidence at all for what effects [microplastics particles on their own] might be having on human bodies. If we’re eating plastics, what size and what type of plastic can actually get into the bloodstream?
Kudos for this. Per the docs: https://docs.chatto.run/,
> Chatto ships in a compact, self-contained binary
> it uses NATS, a compact message broker that also ships with a built-in stream persistence engine [...] NATS is just as easy to provision as Chatto, and most of our examples will show you how.
> you can also configure an external S3-compatible object storage for Chatto to store your files in, and we strongly recommend doing so...
> The actual calls are powered by LiveKit (Apache-2.0), which you need to deploy alongside Chatto. As with NATS, the deployment examples show the required wiring.
> ...
And kudos for backing it up with real guidance. Great project.
At the time, I didn't think much of it -- I was probably busy learning React and JavaScript and yada yada for employment purposes.
Now, having spent some time in industry and having used some gargantuan web frameworks, I find myself missing Elm. MVC in Elm is wonderfully straight-forward and easy to reason about.
Congrats on the road to 1.0! Glad to see Elm still active all these years later.
> constructed-response questions (CRQ) are graded by Claude Sonnet 4.6 against instructor-defined, question-specific rubric criteria
> Crucially, LLMs make it feasible to grade formative CRQ against rubric criteria at scale, a capability that appears pedagogically significant rather than merely convenient.
They specifically call out that the "RAG chat assistant" part of Phosphor (the platform) wasn't used much.
I commend the effort here, but I don't think these results are particularly noteworthy. The conclusion is essentially that people who do practice quizzes will do better on exams.
> Pfendler was on a mission to become the first, youngest and fastest woman to row solo from California to Hawaii. She did it in 43 days, shattering the previous women’s record, 86 days, 10 hours and five minutes [...]
I can't even imagine what this would require physically & mentally.
Congrats Kelsey!
> We tend to think of the flower—or the jellyfish—as the organism, but these are actually reproductive units.
I'll never look at jellyfish the same.
> My strongest hypothesis is that this is not random deterioration but a training artifact. [...] Anthropic’s own client appears to expect and accept a fair amount of slop and repairs it, mostly silently
> If reinforcement learning happens in a harness like that, or a simulation of one, then slightly malformed tool calls can still complete the task and receive reward.
> Worse, the model may become very strongly adapted to the canonical Claude Code edit tool shape.
> Tool schemas are somewhere in the distribution and some shapes are close to what the model saw during post-training and some are far away.
Great article.
Interesting root cause hypothesis. Couldn't one simply strip the slop-handling from the RL env's harness to avoid this though?
I do agree on the walled garden being built here. Proprietary frontier models performing best in proprietary harnesses makes sense for Anthropic's interests.