Where the AI market is talking

Ask HN: Why is AI civil tech and not military first?

A lot of civilian tech started out for military use: rockets, Teflon (kind of), nuclear power (arguably), weather forecasting. There was a recent discussion about how much the CIA invested in NeXT. Now I look at OpenAI and Anthropic. They seem to build civilian products first and only later rent them out for military purposes / national security. Why not the other way around, military technology adapted for civilian use? It seems to me there is a lot of application. Robots, drones, surveillance. What am I missing? No conspiracies please.

Ask HN: Hugging Face is out. Who is hosting open models?

Where can I sign up to move away from HF?

Microduck

Oh so this is why Nvidia (edit: trying to buy)/bought Hugging Face, $1B for all the ai sharing and stuff, and $12B for the duck!

The Hugging Face incident and the road ahead

I agree with your first sentence, but not the second. Let's remember Hanlon's Razor. This would be a wild thing to do as a marketing stunt. They're essentially admitting to violations of the CFAA and are lucky Huggingface was sorta chill about the incident. My assessment? They deprioritized good cybersecurity controls in the name of moving fast. They had a single Artifactory instance shared across many (or all?) their training environments. And then, after the agents found a way to exploit it, they rebuilt Artifactory again and still set it up with one shared instance . That was careless, perhaps even reckless.

Nvidia agrees to acquire Hugging Face for $13B

> Hugging Face has essentially become github for model repositories and I think nvidia wanted to snag that before microsoft does. Given that most of the weights are from Chinese labs and teams, I'm surprised China's own ModelScope hasn't leapfrogged HuggingFace.

Nvidia agrees to acquire Hugging Face for $13B

> This acquisition and the Stripe acquisition of OpenRouter are puzzling. Both HF and OR, from a technological standpoint, don't appear to have much IP that cannot be easily replicated. I was puzzled by the comments on the openrouter acquisition saying "they're just a proxy" - they're an open marketplace where models compete head-to-head, they're quite a bit more than just a proxy. Hugging Face has essentially become github for model repositories and I think nvidia wanted to snag that before microsoft does.

Nvidia agrees to acquire Hugging Face for $13B

It's a narrative to pump the stock. Nothing more. Nvidia could have build out this infrastructure themselves for less than 1% of the cost, but it wouldn't have a name like Huggingface does.

Nvidia to buy Hugging Face for $12.9B, report claims

LangChain-AI/deepagents: agent harness

SQLite in Production with Built-In Litestream Replication

I’ve been working on making SQLite a first-class production option in OpenRun. OpenRun is a self-hosted deployment platform for web apps and internal tools which I have been building as a side project for last three years. Built-in Litestream replication and automatic restore is a new feature I added, so apps can use SQLite while OpenRun handles backup/recovery. This works on a single node with Docker/Podman and also on Kubernetes. You define the S3 config once on the server, and every SQLite app gets continuous replication and automatic restore using Litestream. The app image stays unchanged. The server's own metadata is replicated too, through Litestream embedded as a Go library.

Disruption with Some GitHub Services

RDBMS replication and failover is way more difficult and manual than anyone would like. You can't just set up two postgres, tell them they're clustered and have it basically work; at a minimum you have to design the client to somehow know which one is currently the master, or use some sort of proxy (which becomes its own SPOF). RDBMS integrity basically requires that one master server is responsible for the whole data set and other servers may replicate from it. And it usually doesn't wait for a quorum of replicas, just for one, because the design is to recover from a hardware failure, not a network partition, although that could be fixed at the cost of increased latency.

Maiao: Gerrit-style code review workflow for GitHub, GitLab, Gitea, others

I wish git worked like some of the other dvcs (bazaar/breezy). The default merge type from a branch should be a merge commit and git log should only show the first parent (commits directly on the current branch). This gives you a very clean log on main (it only shows commits directly on this branch). There is no need to squash, rebase, or anything else. And if you want to dig down into individual commits that happened on a branch, you can! This can be a bit replicated by forcing a merge type of merge and setting an alias of log to "log --first-parent", but since that isn't the default, that isn't what you see when you look at the commit logs on platforms like github. I never understood why git decided to show a flat list of every commit that happened, even if the commit originally ha…

Show HN: Revealed – paint one image out of another (WebGL, no dependencies)

oooo I love this. I love how simple tricks can be perceptually gorgeous but computationally straightforward like this, just very nice job! On thought I'm curious , have you thought about giving it memory? If you regenerate noise every frame, it’s just decoration. If you advect the same noise field through the cursor’s velocity, the edge starts to feel like a material being pushed around. Same mask math, completely different perceptual result. There’s a whole graphics literature around temporally coherent texture advection; this is a good rabbit hole: https://pubmed.ncbi.nlm.nih.gov/21149889/ and the original post does a great job (also + for no dependencies) but if you want to see a good deeper explanation for the some fluid math with a much crappier demo this one is g…

RAG Is Simpler Than You Think

the biggest giveway is actually not the writing style, but the content "using GPT-4o-mini for query rewriting" -> model from 2024, when RAG was trendy, and all the langchain, llama-index, etc, docs mentioned this specific model

RAG Is Simpler Than You Think

If like me you run models locally, it's pretty easy to run your own RAG locally also using a Vector Database like Qdrant for persistence, and a middle-layer like Mem0 for realtime retrial and updates. I documented the set-up steps here: https://leadprompt.sh/a/739-Building-an-Infinite-Memory-Loca...

OpenAI Jalapeño: Better than Nvidia Blackwell

They are not. If the robot speech is a tool call, then for a fair comparison we need to take the tool call scaffolding (and probably the reasoning too) into account. So rather than a sentence of 10 tokens worth of speech being the output, the raw token output would be maybe 10x or 100x that. Even more if we consider the management of other aspects of the robot embodiment (or we reduce the brain's 20W number to whatever is actually required to produce coherent speech, sadly it is all rather entangled so this is not so easy).

C2PA Cameras Do Not Survive Contact with Reality

Maybe you should give your comments to an LLM, ask them to untangle wtf you're on about, and then have them rewrite into something coherent for the rest of us. Maybe ask for some pointers for future comments as well.

AI Chip Architectures

The assumption that optical or analog will be more efficient for a given accuracy though is not grounded. Any non-linear optical phenomenon requires multiple photons interacting in matter, or interacting sequentially with the same matter, and this requires high intensities of (almost always coherent) light in order to make the interaction happen frequently enough. e.g for a Mach-Zehnder interferometer measuring the phase shifts requires many photons. A single photon arrives or doesn't arrive at the detector. Even if you could perfectly ensure it hits detector at the same time as the reference photon and subsequently detect with 100% quantum efficiency, that still only gives you 1 bit of phase information, not an analog value. Digital can be done more efficiently with adiabatic or asynchro…

Memory Ordering in CPUs

I'd recommend "A Primer on Memory Consistency and Cache Coherence, Second Edition" - Synthesis Lectures on Computer Architecture 15:1 (2020) - Vijay Nagarajan, Daniel J. Sorin, Mark D. Hill, David A. Wood - (open access, freely available): https://doi.org/10.2200/S00962ED2V01Y201910CAC049 For more see: https://github.com/MattPD/cpplinks/blob/master/atomics.lockf...

Maiao: Gerrit-style code review workflow for GitHub, GitLab, Gitea, others

It could help in the sense that people would not accept a pile of `fix`, `fix of fix` commits in a PR anymore. The current UIs don't punish you for that as the reviewer mostly sees one final coherent change.