Harry Marr

Introducing Claude Opus 4.6. Our smartest model got an upgrade. Opus 4.6 plans more carefully, sustains agentic tasks for longer, operates reliably in massive codebases, and catches its own mistakes. It’s also our first Opus-class model with 1M token context in beta.

Opus 4.5 is a very good model, in nearly every sense we know how to measure. I’m also confident that it’s the model that we understand best as of its launch day: The system card includes 150 pages of research results, 50 of them on alignment.

It’s a good’un - really excited to see this ship. I’ve had lots of moments of joy watching it crush problems I thought it wouldn’t be able to handle. It’s not only smarter, but also more token-efficient *and* 1/3 the cost per token compared to the previous Opus.

Big day - announcing @arcjet's Series A + our new local AI security model 🚀 An opt-in AI layer that runs expert security analysis for every request, entirely locally. 🏡 Accurate detection is the hardest part of security. 🎯 Legacy network-edge tools see packets, not users or logic. Real context lives in your code - where better decisions can actually be made. 🤖 That’s why we built Arcjet’s first AI security model. 🛡️ It runs inference locally in milliseconds, right inside your request handlers. Adds an extra layer to your defenses so you can ship faster and safer. 🍰 Arcjet now protects 500+ production apps used by 1,000+ developers - stopping bots, scrapers, spam, and fake accounts. 🕷️ So we’ve raised an $8.3M Series A led by @pluralplatform, bringing total funding to $12M. Also participating: @a16z, @seedcamp, @feross, and @jeffiel 💸 I'm excited to work alongside a small but exceptional team building the security platform that ships with your code. 🚀

Introducing Claude Sonnet 4.5—the best coding model in the world. It's the strongest model for building complex agents. It's the best model at using computers. And it shows substantial gains on tests of reasoning and math.

If you're interested in AI safety research, I can't think of a better way to spend a few months

⚡️ We've just went live on ProductHunt with @lingodotdev! From a @Cornell AI OSS hackathon to first customers, @ycombinator, @TechCrunch, $4.2M Seed round, and @ProductHunt today! Come see what we've been building, and support us! (link below)

@davidmytton @arcjet @zanelackey @a16z @seedcamp 🚀

I’m excited to announce @arcjet has raised $3.6m seed funding to build the future of developer security, led by @zanelackey at @a16z! Also participating in the round are @seedcamp and a roster of great angels. Although deploying code has become simpler, production security remains too difficult. Developers need a set of tools that solve real problems in the way they’re used to - with code. Arcjet helps developers deal with: - Spam & fraudulent signups - Unwanted bots scraping content - Enforcing dynamic, per user rate limits. - Scanners and attacks like SQLi, XSS, etc Embedding security rules into the application also means they can be tested. If you’ve ever turned on a new security tool then you know the pain of suddenly breaking production because you couldn’t test it locally or on staging! For the last few years I’ve been playing with devtools all day every day for the @consoledotdev newsletter, so I’m excited to be working on bringing a focus on developer experience to security tooling! More details on the Arcjet blog (link below) @arupchak @charlesfitz @tweetsbycolin @duncanjennings @geoffbelknap @harrymarr @ianlivingstone @mxstbr @stopman @dessaigne @nitayj @synopsi @rishabhkaul @arcurn @swyx @t3dotgg @thomaspaulmann

Great to see Quotient launch — congrats @julianeagu and team! Worth checking out if you're building LLM-powered products

@kaythaney yeah at first i just thought it was the folks above us using their vitamix, but then we got some proper creaking and shaking... how about you?

@harrymarr you all ok?

I had no idea that earthquakes were part of the deal when I moved to brooklyn

@MTsireud I tried a third party extension a while back but it didn’t work. Maybe I should try again - I’ve been thinking “just one more month then _obviously_ this’ll be a native feature”, for about 16 months at this point!

@harrymarr There’s been a few openAI tools based on the use case. Have you tried them? Still a good question why google hasnt

A =gemini(prompt, range) function in Google Sheets would be so damn useful. I'd gladly pay for the tokens. It'd be trivial to build and could only help AI revenue. Seems like a no brainer. Curious why Google still hasn't shipped it

@gilbert @AcquiredFM RenTech?

When next week's @AcquiredFM drops, some percent of listeners will go "omg YES" and the others (majority?) will never have heard of the company at all.

@peterejhamilton @incident_io Congrats!! 🥂

We’ve just launched @incident_io On-call and I couldn’t be more excited! 🎉 Our goal was “so good, you’ll break things on purpose” and we’ve poured our hearts and souls into this one. I think it shows ❤️

@zeeg For OpenAI’s models, response_format: { type: “json_object” } should do the trick. For mistral etc, some providers (together, anyscale) have also implemented the same feature and even let you provide a schema: together.ai/blog/function-…

Whats the most reliably mechanism you've found for getting structured output from LLMs? ChatGPT's models were probably 90-95% reliable for JSON, but that error margin is even higher for some others.

I went down a bit of a rabbit hole while trying to predict how many tokens an OpenAI prompt would use... hmarr.com/blog/counting-…

Today, Dependabot will cut false positives and reduce alert fatigue substantially -- to the tune of 15% of all npm alerts. github.blog/2023-05-02-dep…

But don't worry, if you do show it some love, it'll forgive you and come bounding back as eager as ever 🤖🤸

Dependabot's been taking some lessons on human social cues, so if you keep ignoring it, poor lil Dependabot will now take the hint and stop pestering you twitter.com/github/status/…