just_human

Interesting new time series forecasting model from google research that can adapt using examples at inference time (like providing examples in the context prompt for an llm).

TL;DR: Protesters are conducting hunger strikes outside Anthropic (San Francisco) and Google DeepMind (London) offices to demand these companies stop racing toward artificial general intelligence (AGI).

This seems like a premature reaction. The Trump administration has been known to anchor to extreme positions as a negotiation tactic.

Given the complexity of forced platform migrations (user data transfer, algorithm preservation, creator monetization continuity), and the technical/legal hurdles involved, I suspect we're seeing opening moves in a broader negotiation rather than a final outcome.

Let's all be patient and wait to see how this plays out before assuming users will actually have to migrate to a completely new app with new ownership.

tl;dr: Recent research shows animal communication is more sophisticated than previously thought. Bonobos combine calls in multiple ways to create complex phrases with non-trivial compositionality (like human idioms). Sperm whales have their own "phonetic alphabet" with vowel-like patterns and diphthong-like frequency changes. Japanese tits and chimps show compositional communication.

AI tools from projects like Earth Species Project and CETI are helping decode these patterns and may eventually enable two-way communication with animals. While some uniquely human language features (displacement, productivity, recursion) haven't been conclusively found in animals yet, the gap between human and animal communication appears to be narrowing.

Key finding: Animals can create meanings that can't be determined from individual calls alone - similar to human idioms.

No one knows why the Sumerians (later Babylonians) used base 60, but the leading theory is that 60 is highly divisible for a smaller number, making it much easier to do fractions and square roots by hand.

Also interesting, YAML supported base 60 in earlier versions (not sure why? possibly because easier to represent time?), but it was later dropped: https://en.wikipedia.org/wiki/Sexagesimal?utm_source=chatgpt...

Lots of interesting history here, but most relevant was that regulatory and process changes starting in the 80's made it increasingly expensive to build reactors. As a result, reactor construction companies (notably Westinghouse) went bankrupt and no entity was willing to take financial risk to build new reactors. Western Europe is a different story, where political parties aggressively shutdown healthy nuclear plants and passed laws preventing new nuclear.

Much of this regulation and process overhead is now being rolled back in the US (by both political parties) and Europe is slowly coming around to allowing new nuclear. NuScale is one of many next gen companies (I hope they're all successful), but the traditional large reactors are also great and can be built cost effectively.

Recommend starting here: https://world-nuclear.org/information-library/

The problem with nuclear energy is not the availability or the cost of the fuel but the capital cost of the reactor and the high level of financial and operational risk involved with the construction.

Yes, in US and western Europe it's been practically impossible to build new reactors since the 90's for capex and regulatory reasons (both are related). However, we used to be able to build reactors significantly cheaper and faster and I'd argue we're on the path to do it again later this decade. There's no technical reason we can't solve this problem: there's bipartisan support for nuclear, willing financial backers, and no demand shortage. We're going to see 100+ gigawatts of new nuclear in the western world in the next 20 years.

This is somewhat of a clickbait headline. It always bothers me when market analysts are quoted with such definitive statements like "oil will next year fall to between $53 and $56 per barrel." There are so many dynamics to commodity markets - not just supply and demand, but also how market participants will react to an imbalance (for example, will shale producers cut production if oil prices fall).

"We’ve found and resolved two issues that were affecting quality in some Claude responses. We are continuing to monitor for any ongoing quality issues.

We're grateful to the detailed community reports that helped us identify and isolate these bugs."

“… three agents: The first one develops the code based on a requirement, the second one reviews the code and the third one does the git handling (create a branch, checkout, commit)” - this seems to be an emerging pattern to have multiple agents (and eventually agents managing agents).

This is a big deal, not because Microsoft wants to build reactors but because it highlights the real bottleneck: nuclear fuel. There’s already a growing uranium deficit, conversion and enrichment capacity are thin and geopolitically fragile, and next-gen reactors need HALEU, which barely exists today. Building new reactors is the easy part — scaling the fuel supply chain takes years.

Interesting, even with a VPN on mobile safari on an iPhone over a carrier connection I get a uniqueness score of 100%. This is a neat tool, but I'm skeptical of its accuracy. I've run similar tests of uniqueness in the past and this just isn't accurate.

Having worked in a (very) privacy-sensitive environment, the quality of the hosted foundation models are still vastly superior to any open weight model for practical tasks. The foundation model companies (OpenAI, Anthropic, etc) are willing to sign deals with enterprises that offer reasonable protections and keep sensitive data secure, so I don't think privacy or security is a reason why enterprises would shift to open weight models.

That said, I think there is a lot of adoption of open weight for cost-sensitive features built into applications. But i'd argue this is due cost, not privacy.

This is a cool idea! I think you need to find a way to seed more photos (are there any public archives by country/state you could use?), but I could see this being useful when traveling as a way to understand the history of a place.

Somewhat related, but made me remember Color Labs (https://en.wikipedia.org/wiki/Color_Labs) which tried to build a social network where you could see photos taken by other people in your location.

Totally agree solar is awesome, but I always come back to the storage and grid reliability issue. That’s where I get frustrated. If nuclear were easier and faster to build, and built at the costs we see in places like South Korea or China, it feels like it would be the silver bullet for exactly this problem.

tl;dr: In the “ping pong ball paradox” you keep adding balls and removing one, faster and faster, finishing in a minute. Intuitively you’d expect infinitely many balls left, but depending on how you number/remove them, you can end with none (or any number you want).

tl;dr - inventing the wheel wasn’t the hard part, it was inventing the structure to connect wheels to a static system.

I think this is true for a lot of other inventions too. What we usually point to as the invention is really just the surface of a whole system that had to come together. The iPod, for example, wasn’t just a music player. It only worked because mini hard drives, batteries, mp3 compression, etc. The thing we remember as the invention is really the moment when all the pieces finally clicked.

Love this quote from Shizuo Kakutani to describe Polya's result: "A drunk man will find his way home, but a drunk bird may get lost forever."

Link should have been to https://en.wikipedia.org/wiki/Random_walk#Higher_dimensions

Random walks in greater than 2 dimensions are different because you're not guaranteed to return back to the starting point.

probably these companies wanted something quick to get running I've been wondering about that. Assuming that's the case, it seems like eventually CLI will be replaced with a nice UI-based editor.

I still find value in the code editor. I often switch between manually editing code and letting Claude Code run free on smaller tasks, but I haven't gotten close to the point where it can write 100% of the code.

"A drunk man will find his way home, but a drunk bird may get lost forever." - Shizuo Kakutani

I've been using it daily and haven't noticed a change in quality (granted, on a smaller codebase for a single project).

MIT has an amazing history of pranks and hacks dating back to the 1930's.

Congrats on the launch and looks interesting! I love how easy it is to combine local code "tools" with remote mcp servers. The marketplace looks promising, but would be helpful to have some curation as many of the servers don't have descriptions and link to private github repo's. Neat vision and look forward to trying this.

I thought the same thing when I saw this! Great tool but clearly generated by LLM (Claude code?). It’s interesting that it’s so obvious. I wonder if human taste will evolve to dislike styles that are LLM generated.

Awesome! I’ll post it to HN once I have the repo up and the code is in a halfway decent state. Look forward to your feedback!

Not yet, but hope to have something up in September! It’s unfortunately not most compatible - I thought about that but didn’t see a lot of value and there were some downsides like re-implementing an encryption layer that doesn’t make sense if you use WebRTC. Just curious, what’s your use case for mosh compatibility?

Hah, I’ll aspire to do better!