Where the AI market is talking

The Color of White Light

It's a convention. And at one time it was useful as the thing that could be replicated anywhere. Also, things that are white can be judged by the amount of light they reflect or scatter as a percentage, independent of the source of illumination. For instance, among familiar objects, a common white ceramic bathroom tile comes pretty close to pure white.

Maximizing the value of your Claude Code sessions

Considering that I'm mostly unable to reach the limit of my X5 subscription and we have 1M context, it is a guide to maximize Anthropic PnL pre-IPO

Show HN: Visually Precise AI Tutoring on iOS

I trained a model to do perspective correction and now I've got an automated visually coherent AI tutor working from a propped up iPhone. Would love to discuss if anyone has questions! Previous iterations of this were purely on MacOS. This one is on iOS. Link to both is below: https://apps.apple.com/ca/app/knowable-the-ai-tutor/id676378...

Working with AI Feels More Like Leadership Than Coding

I think this comparison holds up. I've noticed that a lot of people I know who get really good results from LLMs and agents are people with significant people management experience. Companies like Anthropic seem to understand that too. It's impressive how many CTOs and CEOs Anthropic have hired for individual contributor positions, which I think is because those leadership skills transfer surprisingly well to working with agents. Of course, managing agents is massively easier than managing humans! You don't have to consider the agent's own desires, goals, opinions, or emotional state when telling them what to do. Humans have agency; agents (despite the name) do not.

Qwen 3.8 27B

We maintain German Langauge index as no one publishes or reruns these sepeartly. Qwen 3.8 27B is a small improvement with some regressions in our benchmarks not a huge jump like benchmarks listed. https://dach.peerbench.ai/compare?models=qwen%2Fqwen3.8-27b,... German language has never been a big focus for asian models but they still outperform Gemma models https://dach.peerbench.ai/compare?models=openai%2FQwen%2FQwe... So in production we have been using Gemini Flash Lite as primary and fall back to Qwen when gemini servers are overloaded or just giving us 429

Suspecting court of using AI, man injected prompts in filings to try to win case

It's a court of law after all, not Anthropic/OpenAI. A bit more credible.

Ask HN: How do you keep up with HN these days?

Or even title keyword. Block [AI, Claude, Anthropic, OpenAI, Codex, Gemini, Grok, DeepSeek, Qwen] and enjoy a clean, pleasant HN.

OpenAI slows release of Astra model citing cyber capabilities

I turned my RSS feeds into an e-ink newspaper to stop reading on my phone

I did this for a while when I was learning Spanish! I used Calibre's CLI on my Raspi which generated the epubs from RSS feeds I'd chosen (Spanish-language media like El País, ABC — I speak in the Peninsular variety) and sent it to my Kindle via e-mail. It did this every morning. Can't stress how much this has helped me with the language. Unfortunately, my Rapsi died some months ago and I haven't gotten around to somehow replicate this workflow without it.

Training AI Scientists to Replicate Research

Google is making private AI practical with homomorphic encryption

I think the term information is misleading here. In any encrypted message there is the original information, it must be because we are able to decrypt it. So from an information theory point of view that the server can "distinguished my ciphertext enough to produce a coherent reply" is always possible. The whole point of encryption is that we have Math that allows the identification of that information only under presence of the (or at least some kind of) secret. The price that HE is paying is essentially "have to run every possible contributing computation", which means it has to scale extremely badly for anything real world complex

NSA and IETF, Part 9

> Let's keep the thread coherent: the original claim How do you save your poisoned wine? A hybrid with 1024 is made less trustworthy if the NSA pushes 1024 alone, since then we know that they want customers to use 1024 alone, which is what they would want if it was weak. But they know that we would know that, so if they really want to help us they should withdraw the draft. If it was strong but we know why, they shouldn't want to make us doubt ourselves. If it is strong (and 512 and 768 are not) they can't tell us, and can only subtly point to their own double encryption and security level documents. The only move that can cover all the cases is a hybrid with 1024, so this draft is a bad standard.

NSA and IETF, Part 9

There is nobody at IETF saying you shouldn't use a hybrid! In fact, it's the exact opposite: hybrid ECDH/MLKEM is a standards-track RFC, and the proposed pure-MLKEM RFC is not, nor is it "Recommended" (in IETF parlance). Let's keep the thread coherent: the original claim, by cryptographer 'cassonmars, is that the issue here is NSA pushing bad standards . It's not "hybrid vs. pure", which is a non-issue. All I asked for was a plausible story about how NSA might have pushed a bad PQC standard.

Yes, Code Review Is Still Necessary

This statement vastly understates the difficulty of putting together a coherent written text at length, but nevertheless its central message is true. There is still a big difference between any text generated by a human vs. an LLM.

Qwen 3.8 27B

The Hugging Face page ( https://huggingface.co/Qwen/Qwen3.8-27B ) shows that Featherless supports the model (and includes the inference widget to try it out directly on the page)

Why does Opus 5 feel worse to work with?

It's not necessarily anthropomorphizing, but simply anchoring. CC learns quickly "this is a session where the user wants to make key decisions". Alas, it is not very good at identifying what constitutes a "key decision", so it keeps asking about all kinds of useless stuff. For that reason I exit session quickly when I can. It used to be that the context of a session is very valuable, because it was so hard to get CC there, but now, this isn't the case anymore, so I only hold onto sessions when there is really hairy stuff that I know would be hard to replicate. I think the whole notion of full automation (long-horizon, subagent swarms, single shot prompting) to have CC build you the whole thing is a pipe dream. CC cannot even write a single doc consistently well. It is excellent at impleme…

Going Dark, and the era of law enforcement hacking

Yes it is, it hacked Artifactory to get to Hugging Face.

Soup Raiders goes native: What you gain by building your own game engine

The Bethesda Morrowind-onward RPGs had their engine roots in Gamebyro, which was fairly widely used by other developers as well. That said I think there's a question about where the value comes from, as their customization to it 25+ years ago and what it enables their game/toolkit to do has clearly given them a 'special sauce' that hasn't been replicated. Also upthread Source was mentioned which has id's Quakeworld/Q2 engine as an ancestor, via HL1 GoldSrc.

Show HN: I built a Claude Code plugin to query 10.6M earnings-call embeddings

Hi HN, I've been using LLMs for one-off market questions against a database of earnings transcripts since ~2025. In tech years that's a long time so as a Solo dev I built a platform around it and have been adding miscellaneous other financial tools and integrations. Backend infrastructure is super boring! PostgreSQL & ElasticSearch, with the data appropriately sharded and replicated across two regions. Some example requests: "What did NVIDIA say about data-center demand on its last call?" "Brief me on NVDA, AMD and AVGO before each market open" "Notify me when AAPL moves more than 5% in a day." The plugin is dependency-free Python and MIT licensed. The corpus, vector search and scheduled-agent service are hosted, so using them requires a free FN2 account. I’m particularly interested in wh…

Everything is about to "go dark"

I think that one of the reasons they (frontier companies and the gov) will be putting so much effort into curtailing bugs and vulnerabilities is to limit the blast radius of future AI models. Imagine with the new Sol Ultrafast, they could have pwned Hugging Face in 6 hours and not 4 days (IIRC). It also seems likely to me that the US Gov. probably already has routine mechanisms for compelling targeted software updates for persons of interest, so I'm not sure that a more formalized backdoor than automatic updates is going to be surfaced in the mainstream, unless that is avenue is also cut down somehow.