There's also GrokAI spewing out CSAM.

There was a list of bad AI writing patterns going around the other day; I can't find it now. Anyone recall it?

The war in Iran was started by a democratically elected president. Not a rogue AI.

ChatGPT was pretty unequivocal that threatening to bomb civilian infrastructure can be a war crime. I wonder if you asked OpenAI leadership whether you'd get the same answer? Journalists if your interviewing OpenAI leaders about their industrial policy please ask. chatgpt.com/share/69d51b...

Google Docs/Sheets/Notion/Word/Power Point. These are still the dominant artifacts produced when using AI to answer business questions. They aren't AI native formats. In a post Claude/Codex world documents have two distinct audiences. 1. Humans 2. Ais

Code mode seems to be working in runmeweb. This should fix a bunch of bugs with the AI incorrectly manipulating the notebook. It should also make it easier to extend the AIs abilities just building out the JS libraries rather than needing to add additional tool calls. @sourishkrout.com

If we want to safely sandbox AI actions do we need a low level interface like WASI which we can use to enforce policy and then build rich libraries on top of that interface?

@hamel.bsky.social @sourishkrout.com I think most OSS notebooks (jupyter , marimo) have asymmetric experiences for creators and readers of the notebook. Pre codex/Claude this was fine because most readers didn't have the know how to read/modify the underlying SQL/pandas code.

What is the sharing/publishing story with nteract? If you use nteract to create a report do readers install nteract? Do you use binder or commuter? nteract.io/blog/nteract...

references to Claude code as an operating system; codex plugins; I hope we're not about to usher in a new duopoly of rent seeking platforms. If your not one of the big three I think the move should be to try to commoditize AI.

If your using code/Claude to answer business questions (e.g. how well are you doing selling widgets to millennials) are you producing shareable artifacts answering that question and if so what is the sharing story? Google/notion/word doc? PDf/HTML uploaded to slack?

Is anyone working on using any of V8 isolates, WASM/WASI to create sandboxes for agents? There are two problems I'm interested in 1. Running agent harnesses in the browser 2. Supporting Google Drive as a filesystem Cloudflare blogged about using V8 isolates; anyone adapting that to the browser?

Who's got a good recipe for agentically coding web apps? Specifically producing videos to verify features are working as expected? I've been using @vercel.com 's agent browser. Screencasts vary wildly when running in headed vs headless. Is it better to control recording explicitly?

Does anyone know if Google OAuth requires a client_secret for the PKCE flow? If I don't include a client_secret I get a token exchange fails with 400 client_secret is missing. Is there some user magic words to trigger a PKCE flow without requiring a client secret? My client is for web.

For me codex (app + 5.3) is an inflection point. I have enough confidence in the AIs coding ability to focus more on design and testing. Concretely; iterating with the AI on design docs and CUJs. CI generates videos of CUJ walkthroughs making it easy to verify tests do what they claim.

AI could be the tower of babel for tech companies. Before AI you needed to pool human capital to build something noteworthy. This was a strong incentive to pay the high cost of consensus building. With AI you can get much further by oneself. But that only postponed the reckoning.

As I watch what's happening in Minneapolis and around the country ; to me the immediate risk of AI is that it's consolidating power and wealth in the hands of tech executives that willingly endorse what's happening as long as they get their GPUs.

To all my new England friends. It's 61 degrees and sunny in the bay.

Twitter died a long time ago and X isn't worth it. Before I made the jump, my ego was holding me back. I was too concerned with vanity metrics. I finally realized if you're posting stuff people want to follow, then they'll follow you, even when you leave. www.siliconrepublic.com/enterprise/k...

Police Repeatedly Shoot Tim Cook After Mistaking iPhone For Gun https://theonion.com/police-repeatedly-shoot-tim-cook-after-mistaking-iphone-1824184361/

One of the best things I’ve read in a while:

I work in tech. I emailed my leadership letting them know I hope they acknowledge the events in Minneapolis and take a position. Small as it is; there was a time when I wouldn't have had to ask. I don't know how but I want to try to hold tech accountable.

This is ok tho

I think the market keeps wanting to believe things haven't changed but they have. We are very clearly in an era where the president is willing to go after anyone, any country or any company that displeases him. The volatility should scare most investors. www.nytimes.com/2026/01/21/b...

That didn't take long

I don't understand the argument that acquiring Greenland is vital to our national security. I can't think of anything more dangerous and a bigger gift to Russia and China than the fracturing of NATO. How are these clowns in charge? www.nytimes.com/2026/01/20/u...

If we think aggressively pursuing Greenland and threatening NATO is bonkers and undermines US security and economic interests; what do we do? How do we encourage Congress to fulfill its constitutional responsibility and restrain the president?

The post is a good explanation of memecoins and the attempt to boost their value by tying them to OSS projects

What options are there for replacing my LinkedIn page with a resume hosted on PDS/AtProto 1. Do we have an existing schema/proto for storing it as structured data 2. Is there a webapp that can be used to edit mine and view others

Has the tech industry crossed over from optimism to toxic positivity? There was a time when events like what's happening in Minneapolis would have elicited at the least internal comms acknowledging how troubling that was. Is that happening? I don't think so.

one reason we keep doing coding interviews even though they no longer make sense is because we know how to do them using coderpad. I think coderpad could help advance the field by supporting new styles of questions. E.g here's a large code base use AI to figure out X works

Finally got a round to deleting my X account. Exporting my data was surprisingly easy. I assume that is a testament to the great engineering at Twitter pre Elon.

We need a symbol like TM to indicate "I don't really know what this means either so your best bet is to ask AI." Here's an example using AA as the symbol. "We use chz blueprints[AA] to manage the definition of our eval jobs."

It's taken my new puppy less than 24 hours to turn my house into a complete bio hazard zone.

I'm optimistic that AI will commoditize compute and lead to a new wave of apps running on compute you own. Cloud compute has a lot of benefits (e.g. reliability) but managing EC2/GCE was too hard. AI is a powerful new tool compared to SAAS which trades ease of use for loss of control

I wonder how many people aren't trying Codex CLI and Claude Code because they think it's for SWEs and they don't consider themselves SWEs? Is a CLI the right UX for this audience? Does it need to be a Cloud IDE like replit?

Git is so deeply ingrained in software engineering it's hard to remember what a barrier to entry it can't be. In a world where anyone can write code with AI I don't think the answer to "where do I save it" can be create a git repo. The UX needs to be as seamless and intuitive as Google Docs.

My team is looking for a data engineer with expertise in Kafka and data bricks. Everybody I ever worked with at Google on Dataflow or Primer would be a good fit. So if your looking reach out to me.

Software engineering and coding are not the same thing Coding = getting persnickety machines to do what you want with esoteric syntax Software engineering = solving business problems with software. AI will solve coding but not software engineering; that requires customer empathy

Two central competing AI narratives are 1. AI let's you do the same work with less people -> job loss 2. AI let's you solve problems you couldn't before -> create entrepeneurs I think enterprises are structurally geared to the former; i.e layoffs and stock buyback

I've accepted that 1. I'm building Google Colab 2. Codex is enabling this by letting me cosplay a web app developer 3. I'm more than likely tilting at windmils

Does the atproto have good support for non-public data? Could a PDS be used as an alternative to Google Drive? @pfrazee.com

Trump going after Maduro while pardoning Juan Orlando Hernández only makes sense to me if I think of Trump as a chaos agent. I think he treats each moment as an opportunity to win or lose and does whatever feels like winning without any long term planning or strategy.

Skills are very interesting but I think the key to unlocking AI remains creating a UX where training data is a side effect of getting stuff down. In skills the heavy lifting is done by a skill.md file which is handcrafted, training data. leehanchung.github.io/blogs/2025/1...

I'm not on the multi-agent bandwagon yet. I've been using codex-cli to build my own notebook for devops. Codex has been a game changer for learning and building a webapp. But it required a huge amount of steering.

Is it possible to create a first class sharing/ Google Colab like experience in Jupyter/JupyterLab or does the architecture prevent it? Are the frontend, backend and storage so tightly coupled that its difficult to separate them and host the frontend separately.

Software development is unique in that development environments are hermetic and cloneable. We have decades of CI/CD investment towards that goal. This makes development tasks well suited for evals and RL which I think helps explain the rapid progress.

Here's my enormous round-up of everything we learned about LLMs in 2025 - the third in my annual series of reviews of the past twelve months simonwillison.net/2025/Dec/31/... This year it's divided into 26 sections! This is the table of contents:

I'm so sick of subscription pricing. AI might make a difference here because to me it looks like consumption based pricing is the only sensible approach. Intelligence = compute = energy. I don't see how a business selling compute and energy can provide unlimited compute and energy for fixed price.

Anyone have recommendations for books or online courses to learn frontend engineering and webapp design