Where the AI market is talking

Now Anthropic Is Saying Claude Escaped and Hacked Several Companies

When techies hear that “AI escaped a sandbox” we understand that it was able to communicate with the outside. When ordinary people hear “AI escaped and hacked” they think about Mission: Impossible where an amorphous entity has replicated its execution environment and uploaded itself to another server to live there: a worm.

OpenAI's Hack of Hugging Face

Show HN: Kedge – Full-stack cloud with forkable VM snapshots and global SQLite

Built in replicated database and filesystem is pretty rad. Fly leaves this as an exercise for the user. You should license the backend to enterprises that want a cheaper/lighter weight option to kubernetes for the surge of agentically coded internal apps.

Show HN: Kedge – Full-stack cloud with forkable VM snapshots and global SQLite

Hey this is amazing! I see the sqlite db is replicated but how does it handle multiple instances as writers? Curious if the underlying sqlite engine is syzy or if there's other restrictions to be aware of

Anthropic AI Models Hacked Three Companies During Tests

I said this about OpenAI/Hugging Face, and I'll say it again for Anthropic: It's not that I think this is fiction; I'm confident these events actually happened. But I think they were effectively allowed to happen because Anthropic and OpenAI are constantly chasing each other for the narrative of "hugely advanced, maybe almost sentient AI lives here." It reads more like a press release than a security update, and I think that's because it is. I hope this type of marketing backfires and the companies face actual scrutiny and consequences for operating this way.

2x, not 10x: coding with LLMs in 2026

Since I'm on the Product side these days, I don't spend much time delving into the codebase at this point in my career. I can believe it though, as I've done quite a few personal side projects using langchain and various harnesses. The problem I see is not necessarily that LLMs build things badly (although I know that can also be true), its that because people are using LLMs to generate documents, we are then building the wrong thing entirely (whether with an LLM or not, whether badly or not). Removing human thought from the process of development is generally a net negative, IMO.

Hugging Face Incident Initial Post-Mortem

Stacked PRs are now live on GitHub

I don't have experience with the stacked PR workflow and often find myself asking the same questions you are. As far as I can tell, the biggest benefit of stacked PRs over just making a coherent series of commits, is that it might make it easier to start work on your second PR before you merge the first one? With human-in-the-loop coding, that sounds like it could lead to a lot of wasted work if the first PR gets substantial feedback. But with agentic coding, I can imagine how it might be desirable to keep the agent chugging while the first PR is under review. Interested in learning more about it and generally agree that AI is stressing the current review paradigms a lot of us are accustomed to.

2x, not 10x: coding with LLMs in 2026

From my experience, a snippet of code will look good. But LLMs seem to have a hard time getting the big picture and reusing code that is already implemented and ALMOST does what you want vs. rewriting everything from scratch. AI code is extremely difficult to read and follow. It's littered with hundreds of lines of comments and notes, often referring to other notes in other parts of the codebase, and often extremely out-of-date. Claude seems to me to still focus on getting things done vs. doing it properly, so from my experience it would pretty often mess up your whole codebase just to be able to finish its task, vs. stopping and rethinking the approach. After a few passes of that, with duplicate code with no underlying coherent vision, hundred of thousands of lines of documentation writt…

Ubiquitous Language is prompt engineering that humans can read

I very much agree with and relate to the stance this article takes, that a domain model expressed in plain text that is both human readable and agent accessible is a _very_ good idea to keep agent-generated code more coherent than it otherwise would be. I’m less a fan of the described approach of putting the domain knowledge in CLAUDE.md however. My preferred approach, because the domain model should be a discretely and separately managed asset, is to store it in its own file that is referred to by the agents, with a tool that allows me to edit and visualise it for sharing with the humans as well. This tool is the one I built to do that for the core concepts in any domain, stored in plain text for the level playing field with agents and humans: https://thinkingtools.software&#x2…

The Economic Benefit of Refactoring

I think the name of the game is going to be "how do we maximize output while keeping token counts low". For example, finding ways to replace AI workflows, even if minor ones, with scripts/code. Anthropic told me that I have till August 19th to be get my act together becuase they are going to reduce my token count by 50%. lol. I have been abusing my Pro Max allowance and need to start being less wasteful. Articles like this, can help us come up with ideas on how to do it.

Making Postgres queues scale

We are very heavily using Postgres as a queuing system in production for many years already, not just some quiet background tasks but with millions of tasks in the queue at any given moment. I have yet so see any issues with that so I'm always a bit suspicious when people say they had to reach for something else unless you are at a crazy scale. I know it's very hard to compare workloads, but famous recent example: https://openai.com/index/scaling-postgresql/ > It may sound surprising that a single-primary architecture can meet the demands of OpenAI’s scale; however, making this work in practice isn’t simple.

Citadel buys Situational Awareness after steep AI losses

Situational Awareness will be a good case study for whether or not circular deals unravel, or if the AI industry is more resilient than critics fear. Here's why: 1. Situational Awareness was recently leading Fluidstack's $830 million Series A. 2. Fluidstack is helping lead Anthropic's $50 billion data center buildout. 3. Fluidstack is one of TeraWulf's major customers. 4. Some of Terawulf's bonds are tied to Fluidstack's leasing for the Anthropic operation. 5. Google is backstopping the bonds above, should Fluidstack not be able to pay TeraWulf. I heard via a number of sources (e.g., CNBC) that some of Situational Awareness' private deals might not move forward as a result of what's going on. It will be interesting to see if any data centers, leases, etc. become problematic as a result of…

Banks Line Up $15B of Debt for Anthropic with Google Aid

Advancing the price-performance frontier with GPT‑5.6

I pretty strongly disagree about comparing this to Kimi and GLM, 5.2 was a big price hike for Chinese models, and Kimi K3 was a big price hike to that. K3 was within spitting distance of OpenAI pricing (more expensive than short context Terra, less than long context). And that's after months of OpenAI/Anthropic prices going up. Now we have an American lab drastically cutting a price, feels like this is the opposite of that trend.

Advancing the price-performance frontier with GPT‑5.6

I pretty strongly disagree about comparing this to Kimi and GLM, 5.2 was a big price hike for Chinese models, and Kimi K3 was a big price hike to that. K3 was within spitting distance of OpenAI pricing (more expensive than short context Terra, less than long context). And that's after months of OpenAI/Anthropic prices going up. Now we have an American lab drastically cutting a price, feels like this is the opposite of that trend.

UEFA and its national associations will not participate in FIFA competitions

These people looked at OpenAI and went "Hey, I can run corruption like that too!"

Advancing the price-performance frontier with GPT‑5.6

If you optimize program A and manage to wring out a 1% improvement, and I optimize program B and improve performance by 20%, you can see the problem with trying to infer anything from those two numbers. Edit: searching for the story now, further bolstering the point is that was 1% in training time [1], and the openAI claim is 20% in end to end inference cost . This is a bad comparison. [1] https://deepmind.google/blog/alphaevolve-a-gemini-powered-co...

Advancing the price-performance frontier with GPT‑5.6

Friendly reminder that GPT-5.6 in Codex cannot spawn Luna subagents, only Terra [1] To fix this you currently need to make your own copy of the bundled model catalog [2] and opt Luna into MultiAgent V2. 1. https://github.com/openai/codex/issues/32031 2. https://github.com/openai/codex/issues/32031#issuecomment-51...

Hugging Face is being used to easily undress women and children

> Hugging Face is being used to make nonconsensual deepfakes, and the popular open-source AI model repository is doing very little to prevent it. That’s according to a new report published by the European nonprofit AI Forensics, which found that seven out of the top nine image editing models hosted by Hugging Face readily complied with requests to undress women using simple prompts. A, yes, because there wasn't enough censorship in the world. Now we need to invent stupid "think of the children" reasons for locking general purpose tools like open models. Of course the whole point is moot, people wanting to do that, can just keep using whatever open model outside of hugging face, even exchange models privately with other pervs. Once they're out there, they're out there. So, this is just …