Instant nostalgia: https://youtu.be/tQJJ_rhHxIk?si=V8pOVwj3gYkw0xKB&t=188

In March Karpathy described this direction:

   The next step for autoresearch is that it has to be asynchronously massively collaborative for agents (think: SETI@home style).

Seems like Karpathy was largely focused on ML / SWE research rather than the other domains this group is after. Still, hard to imagine they were not influenced by autoresearch.

Andrej, if you're around, please share your thoughts on Discovery Loop.

However, reducing (or rather limiting the increase of) PII leakage and impact of ransomware activities is much closer to day-to-day mainstreet of most people.

Anyone committed to advancing science should care about this regardless of its potential contributions to defense.

https://banagale.com/cinematic-canvas-ai-film-animation.htm

I provide a "how I got to this" up front, but if you want to jump right to the Apocalypto stuff, use this:

https://banagale.com/cinematic-canvas-ai-film-animation.htm#...

And if you want to play with interactive demos of the two animation sequences (the tuning tooling I described in my OP) you can go directly there:

https://banagale.com/cinematic-canvas-workbench-demos

If anyone wants to collaborate on building out the cinematic-canvas-workbench project please email me.

If you have time to try the product as it stands, I would definitely appreciate your feedback in particular, and would be interested in checking out what you've built so far for yourself.

I will publish the API to the database as well.

That page with the time machine may not make it obvious, but the product does have a linux client!

It doesn't have the same app window and summarization of the macos version, but the transcript ingestion engine is efficient and the real value is in leveraging the database it builds using the packaged `/total-recall` or your own use of the api. (which is not yet documented but discoverable)

I am building the windows version now. I have been for the past four days. It uses a shared swift-core with the macOS and Linux versions which has been part of the reason it has taken "so long."

If anyone is on windows (or linux!) and would be willing to try it either of the clients my email is in my profile, I would be grateful.

Also I'm working up a short post with that apocolypto anim now.

That took a fair amount of custom tuning and I had to create a tuning view to get some of the behaviors right.

But it was enough fun that I generalized it to take in ~any scene description from a film. It goes out and gets more detailed descriptions and film stills if available but also takes custom stills if you provide them.

My test scene was the Gauntlet scene from Apocalypto. It is low fidelity but does a pretty amazing sequence with somewhat believable physics of the javelins etc.

Here is the docs page with the vertical takeoff / 88 miles an hour time travel: https://contextify.sh/docs

I can share some of the Apocalypto bit if anyone is interested.

Because long running, well tuned pipelines can deliver on some important goals but be broken by the changes from adopting a new model (ahem opus 5) or rot in some key web automation path.

   The challenge with screen mirroring solutions is that they take over every single pixel in the car, and that’s not the way we see ourselves interacting with our users.

[0] https://news.ycombinator.com/item?id=48769397

Though this is especially helpful for old homes where you have to sort of improvise ethernet wiring.

I normally just use it to consume 4k hdr media though.

It has not lived up to being a utility player device, but so far it is so good at media delivery that is sort of okay.

Comments URL: https://news.ycombinator.com/item?id=49056022

Points: 28

# Comments: 13

Not just in what the models can or might want to do, but how they treat the operators they interact with.

If you look carefully, this card shows the addition of a new benchmark for "condescension" as a character trait.

I think a lot of people would like to see a comparable system card for the unannounced model that escaped openai last week.

What terminal tooling are you using?

During post-training of opus 5, the last few days, opus was a real wreck. I had to swap in gpt 5.6 sol for my orchestrator and enable fast mode (1.5x speed) in order for it to keep up with work and communications from a handful of mostly 5.6 sol agents.

Also because interacting with a slow orchestrator is no fun, even when plenty of work is getting done in parallel in the background.

One of my first jobs out of undergrad was technical writing on EDA tools for Mentor Graphics Calibre product. For many years, I did not appreciate that experience--it seemed orthogonal to what I'd gone to school.

Once I did become a professional SWE, something that took me too long to realize was that writing documentation and tutorials for people and looking after the quality of onboarding materials was almost never rewarded.

I came to see it more of a liability with ~all emphasis being on human code gen.

I listened to an interview with Steve Yegge earlier this year, one of his concerns was that many people just don't type very fast or read very quickly. IIRC, he expressed concern this would continue to be a big barrier to successful use of AI.

I previously commented on annotation for coding agents that an important skill was knowing the beat of a conversation.

I think that extends into critical reading. For example, it is not that unusual for a Washington Post article to put some ~explosive detail deep in the article--this is called "burying the lede."

I see this regularly reading AI output. That is: the AI will lead with the strongest claim it can state cleanly, but not the most explosive implication.

With Claude Code and Codex, you're looking essentially neverending walls of text dense with technical information.

So now you have to have high domain expertise, a good understanding of written communication, a sense for the beat of a conversation (which assists in identifying understatements or overstatements), and the ability to dense text outputs quickly and then highlight critical statements for further exploration.

I'd almost describe part of this work as that of a skilled interviewer.

Glad for all of his contributions, I did enjoy reading his work.

But I think your comment drives at some authentic take on this. Skill with AI is not only crafting iterative prompts the agent will understand, but also very high domain-specific knowledge of what the prompts explore.

One without the other can result in frustration or worse.

The deployment target is either a garage or a home "utility" room where the robot would be mounted to a table or ceiling mounted and intent might be sorting hardware, doing laundry or packing suitcases.

The cobots I'm interested in are Universal Robots UR5, UR10 (you're gushing with cash to experiment) or down to AR4-class hardware, or SO-101 plus lerobot (you're doing a lot of making).

Comments URL: https://news.ycombinator.com/item?id=49010573

Points: 3

# Comments: 0

Because I absolutely feel I made entire SaaS products, cool ones, by hand and they sat on a multitude of packages and infra I but glued together.

For frontend, I tailored templates, and instead put energy and time into making modern frontend toolchains and Django backends work together. This was satisfying to understand and build.

Being at ease with the minute details of Django or latest flags of esbuild felt like knowing the various modes of my Dewalt or Dewalt-colored power tools.

Building with AI is not a passive activity, and doing so well and efficiently, there is a ton to master and it evolved a great deal in the past six months.

So leveraging the tools using custom skills, custom CLIs, etc this is very important and very valuable thing to do. I would classify it as a serious contemporary computer science skill that should be taught in addition to the fundamentals.

Presume now, you're making and leveraging all of the modern bells and whistles of Claude Code and Codex. That is, you're reading the release notes and you have personal tooling so you can switch between them easily.

You can build incredible things with sustainable release workflows and reasonable security and possibly more than what a solo dev's "production quality" of yore.

I know this because I had Fable look at an entirely "hand crafted" SaaS I built over two years and it found about a page of bullets just in p0-p1 that I had not caught in my artisanal best effort. They were real issues, maybe unlikely but still things that I would have fixed if I had known.

But what can not be replaced is *taste*.

You can build all day and night and if it ain't good people won't use it. If you can't describe it in a way that makes sense to people they won't care.

If you build something people don't want you've not really made anything more than we did before we had AI as a tool.

I do think it can still achieve the same level of satisfaction if it is built well. Because it actually does take a lot of skill and knowledge to make something well using agentic programming. This is regardless of whether people want it.