beklein

They mentioned "... and my visualizer tool comes with an MCP server ...", with a link to the rerun project. I guess it makes more sense that his visualizer tool uses rerun...Sorry for the confusion from my side.

A bit off topic, but I absolutely love the little robot on the author's main project's landing page (https://rerun.io/). I normally condemn mouse hijacking, but this implementation will be allowed.

I don't know what kind of experiences and apps this UI/UX mode will enable but it's one of the most innovative and coolest AI model demos I've seen this year.

"In fact there are only two things universities need to change to be perfect at preparing founders: they need to make students feel that starting a startup is something they can do, and they need to encourage them to work on their own projects."

I'd argue that both of these would be highly beneficial for all kinds of students, founders and non-founders alike.

Starting a startup could be generalized for all students to having a dream or goal, taking ownership of it, and making it happen. A startup is just one possible path, pg is probably too specific here, I would agree.

Having more free time to work on my own projects is probably the freedom I missed most at university. I think learning to choose what to work on and making something happen would benefit future workers, teachers, doctors, and engineers just as much as entrepreneurs.

It's a trade-off in terms of time and other resources, since parts of the existing curriculum would have to fade away. But I think it's a good starting point for change and a good direction to move in.

Thanks for sharing your thoughts. One of the best frameworks I have encountered, and one I point people to whenever I can, is the AI Fluency framework by Anthropic, Prof. Joseph Feller (University College Cork), and Prof. Rick Dakan (Ringling College):

https://www.anthropic.com/learn/claude-for-you

This framework provides a more abstract way of thinking when working with AI systems. Especially valuable to me is the pattern of the 4 D's: Delegation, Description, Discernment, and Diligence, which I have now internalized whenever I use AI systems. It helps me understand my own and others' (mis)use of AI systems a bit better.

Thanks for pointing that out!

I actually copied the link from NVIDIA's Technical Blog post:

- https://developer.nvidia.com/blog/nvidia-nemotron-3-5-lightn...

You can also try the model via a free API endpoint from Openrouter, would be interesting to see if it's the BF16 or NVFP4 version:

- https://openrouter.ai/nvidia/nemotron-3.5-lightning:free

I guess the author used a different framework but I like this one: https://animejs.com/

Thanks! I also like the OpenAI Cookbook: https://developers.openai.com/cookbook

Other AI labs also tend to publish examples and cookbooks on GitHub and Hugging Face, so it's always worth keeping an eye on those as well.

Details about the methods can be found here: https://artificialanalysis.ai/methodology/intelligence-bench...

Specifically they use this harness: https://github.com/ArtificialAnalysis/Stirrup

Some other (as in better) sources I found:

- https://huggingface.co/spaces/Soofi-Project/Pretraining-Tech...

- https://arxiv.org/pdf/2607.09424

Thanks for sharing the paper, it seems my intuition on size was bad but also my intuition on max satellite density for LEO was wrong.

To get a better sense of the scale, if you are viewing this app on a 4K display, with the planet measuring about 2,000 pixels across Earth’s diameter is approximately 12,742 km (7,918 miles), so each pixel represents about 6.37 km (3.96 miles).

A Starlink satellite is roughly 6 m (20 ft) wide without its solar panels. This means a one-pixel satellite marker is shown at roughly 1,000 times its true size. So even if this image already looks extremely crowded, the dots are still massively exaggerated. Visually, there would be roughly another factor of 1,000 before the satellites themselves were shown at their true scale—although this does not mean that orbit could easily accommodate 1,000 times more satellites but I guess there is still some space in space.

All AI labs really need to stop using truncated y-axes for benchmark bar charts...

https://mistral.ai/_astro/cm-engish_ZhlvoT.webp?dpl=6a3a94bd...

A good visual explanation of how text diffusion models like DiffusionGemma work: https://newsletter.maartengrootendorst.com/p/a-visual-guide-...

Principles of programming:

1. Break things down into small units

2. Think about sequence

3. Find patterns

4. Focus on the important things

5. Visualize sequences in your mind

Love the silly music and the way they teach, thanks for sharing this!

Also relevant to this is the newest episode of The Lightcone Podcast with Quan Vuong, co-founder of PI and, one of many co-authors of that paper.

Thanks for sharing this here, it's a beautiful video. The images are linked in the description but if anybody is reading this, check: https://www.flickr.com/photos/nasa2explore/

"... the first early version of Claude Mythos Preview was made available for internal use on February 24. In our testing, Claude Mythos Preview demonstrated a striking leap in cyber capabilities relative to prior models, including the ability to autonomously discover and exploit zero-day vulnerabilities in major operating systems and web browsers."

More infos here: https://red.anthropic.com/2026/mythos-preview/

As a big Codex user, with many smaller requests, this one is the highlight: "In Codex, GPT‑5.4 mini is available across the Codex app, CLI, IDE extension and web. It uses only 30% of the GPT‑5.4 quota, letting developers quickly handle simpler coding tasks in Codex for about one-third the cost." + Subagents support will be huge.

Not sure why you think Anthropic has not the same problems? Their version numbers across different model lines jump around too... for Opus we have 4.6, 4.5, 4.1 then we have Sonnet at 4.6, 4.5, and 4.1? No version 4.1 here, and there is Haiku, no 4.6, but 4.5 and no 4.1, no 4 but then we only have old 3.5...

Also their pricing based on 5m/1h cache hits, cash read hits, additional charges for US inference (but only for Opus 4.6 I guess) and optional features such as more context and faster speed for some random multiplier is also complex and actually quiet similar to OpenAI's pricing scheme.

To me it looks like everybody has similar problems and solutions for the same kinds of problems and they just try their best to offer different products and services to their customers.

Perhaps useful, I discovered: https://github.com/agent-infra/sandbox

> All-in-One Sandbox for AI Agents that combines Browser, Shell, File, MCP and VSCode Server in a single Docker container.

Some more info here: https://developers.openai.com/api/docs/models/gpt-realtime-1...

- $4 input, $0.4 cached input, $16 output

- 32,000 context window

- 4,096 max output tokens

- Sep 30, 2024 knowledge cutoff

Love the models, speed, and capabilities. Just sad that they are not getting the publicity and adoption right now, but hopefully in the future.

Sound on!

Song name is: Windowdipper from ꪖꪶꪶ ꪮꪀ ꪗꪖꪶꪶ by Jib Kidder

https://jibkidder.bandcamp.com/track/windowdipper

The end result would be a normal PPT presentation, check https://sli.dev as an easy start, ask Codex/Claude/... to generate the slides using that framework with data from something.md. The interesting part here is generating these otherwise boring slide decks not with PowerPoint itself but with AI coding agents and a master slides, AGENTS.md context. I’ll be showing this to a small group (normally members only) at IPAI in Heilbronn, Germany on 03/03. If you’re in the area and would like to join, feel free to send me a message I will squeeze you in.

Not my normal use-case, but you can always fall back and ask the AI coding agent to generate the diagram as SVG, for blocky but more complex content like your examples it will work well and still is 100% text based, so the AI coding agents or you manually can fix/adjust any issues. An image generation skill is a valid fallback, but in my opinion it's hard to change details (json style image creation prompts are possible but hard to do right) and you won't see changes nicely in the git history. In your use case you can ask the AI coding agent to run a script.js to get the newest dates for the project from a page/API, then it should only update the dates in the roadmap.svg file on slide x with the new data. This way you will automagically have the newest numbers and can track everything within git in one prompt. Save this as a rule in AGENTS.md and run this every month to update your slides with one prompt.

In my AGENTS.md file i have a _rule_ that tells the model to use Apache ECharts, the data comes from the prompt and normally .csv/.json files. Prompt would be like: "After slide 3 add a new content slide that shows a bar chart with data from @data/somefile.csv" ... works great and these charts can be even interactive.

I love this! I use coding agents to generate web-based slide decks where “master slides” are just components, and we already have rules + assets to enforce corporate identity. With content + prompts, it’s straightforward to generate a clean, predefined presentation. What I’d really want on top is an “improv mode”: during the talk, I can branch off based on audience questions or small wording changes, and the system proposes (say) 3 candidate next slides in real time. I pick one, present it, then smoothly merge back into the main deck. Example: if I mention a recent news article / study / paper, it automatically generates a slide that includes a screenshot + a QR code link to the source, then routes me back to the original storyline. With realtime voice + realtime code generation, this could turn the boring old presenter view into something genuinely useful.

Honest question: would a normal CS student, junior, senior, or expert software developer be able to build this kind of project, and in what amount of time?

I am pretty sure everybody agrees that this result is somewhere between slop code that barely works and the pinnacle of AI-assisted compiler technology. But discussions should not be held from the extreme points. Instead, I am looking for a realistic estimation from the HN community about where to place these results in a human context. Since I have no experience with compilers, I would welcome any of your opinions.

Relevant blog post from simonw: https://simonwillison.net/2026/Feb/7/software-factory/