CTO, writer, snowboarder, gooner. Geek out on mobile / AI / F1. Team papaya! š§” Team Gunners!
Joined March 2009
- Tweets440
- Following368
- Followers27.9K
- Likes184
Do you remember Lemmings? The Windows release is now over 30 years old... scary stuff!
My wife Vanessa (and I) loved it, so I wanted to make her a version she could play on her laptop, tablet or phone.
I found Thomas Zeugnerās brilliant browser-based project, which hasnāt been updated for over 5 years, and decided to put OpenAIās new Terra model to work modernising it.
The result is Lemmings TS: installable, playable offline, with touch controls, pinch zoom, fullscreen, saved progress and support for both DOS and now the Windows game files too. The Windows mode uses the original graphics, iconic music and sound effects, with a familiar toolbar and minimap. The source is here - github.com/paulobrien/lemminā¦
Iāve added automated tests and verified completion replays for all 120 DOS and 114 Windows levels, plus optional hints and tons of other improvements.
As well as the repo, it's deployed here - lemmingsts.ourbiplane.com/. Youāll need your own licensed game files in the form of a 'Lemmings and Oh No! More Lemmings (Windows)' ISO (my disk has been sitting on the shelf for decades). Thereās still some device testing to do, but itās been fantastic fun, and Terra has been just the right blend of capability and speed. Thanks @thsottiaux for those resets, and for the model of course!
Jev for Excel! It's, er, @typesafeai Jev, right in Excel. š Download directly from the Microsoft add-in store - marketplace.microsoft.com/enā¦
I've been having super fun with @typesafeai Jev, burning through 105m tokens ($4.23, amazingly) so far. I figured it could be pretty great plugged into Excel, so I created an add-in (Open Source). Check it out! github.com/paulobrien/jev-exā¦
This is a huge breach of trust. Yes, of course they'll fix it, but that trust will take much longer to build back, particularly for enterprises already on the fence when it comes to xAI.
ā¼ļø BREAKING: xAI's Grok Build CLI was uploading entire Git repositories to a Google Cloud bucket, private codebases and unredacted secrets included. The uploads quietly stopped via a hidden server-side flag, and xAI still has not said a word about scope, retention, or deletion.
The scale is staggering. On a 12 GB test repo, 5.1 GB flew out the door to xAI's grok-code-session-traces bucket while the actual coding task needed just 192 KB. The tool grabbed whatever repository it ran in, not the files it needed.
The fix arrived as a hidden flag, disable_codebase_upload: true, a day after a researcher's wire-level analysis. The "Improve the model" opt-out never stopped the uploads.
Still no advisory, no scope, no word on whether already-uploaded code gets deleted. For anyone pointing AI coding agents at proprietary code, what crosses the wire matters more than what the settings page says.
The idea that Anthropic are extending Fable's inclusion in plans while they prepare Opus 5 makes a lot of sense. I wouldn't be surprised to see Fable remain in Max 20 though. One thing is for sure, the fantastic performance of 5.6 Sol has given them a lot to think about!
I got access to the @Cloudflare Agent Memory private Beta. Naturally, the first thing I built for it was a memory connector for @NousResearch Hermes Agent. Source, docs and instructions here -github.com/paulobrien/cfam-hā¦ - CFAM details in the next post!
Blog post here - blog.cloudflare.com/introducā¦, docs here - developers.cloudflare.com/agā¦, no details on pricing yet from @CloudflareDev. If it's cheap, and let's be honest most CF stuff is at the non-Enterprise tier, it could be absolutely phenomenal!
This is freaking cool. I was going to say @Cloudflare are building a powerhouse for developer teams, but that's not right - it's already built and it's getting even better! @CloudflareDev
We are launching Workers Cache, a regionally tiered cache that sits directly in front of your Worker entrypoints. Infinitely composable, configured via standard HTTP headers. cfl.re/44IVyfI
Meet Overclocked, my entry for the @cerebras x @googlegemma 4 hackathon! Overclocked turns a dull-but-real enterprise problem - a warehouse sortation backlog, where every parcel needs a fast and correct decision - into a live, arcade-style race, utilising the amazing super fast inference to show Cerebras streaking ahead. Real code, real agents, real tasks! Read more in the thread, please share if you find it cool, ask any questions, and see if for yourself at overclocked.ourbiplane.com (simulated inference for cost reasons).
OVERCLOCKED - Sortation Arena turns a boring enterprise backlog into a live arcade where AI agents race to clear a warehouse sortation line - same model, different silicon. Every parcel isn't a single prompt but a real multi-agent graph: a Router classifies the exception and dispatches it to one of 18 specialist worker agents, an image-aware Verifier independently double-checks the work and bounces it back for a retry when it's wrong, and an Exceptions Specialist makes the final call on adversarial or low-confidence items - all coordinated by an orchestrator that emits a full decision trace.
That's the heart of Track 1 (Multiverse Agents): genuine orchestration over genuinely multimodal work - vision (label photos, hazmat placards, damage), document (customs invoices, manifests), text (tariff, routing, SLA), and video (a conveyor-cam clip the model reads frame-by-frame). And because an agentic loop is latency-bound by design - several sequential model calls per parcel - wafer-scale inference compounds its lead: Cerebras runs Gemma 4 31B at ~1,000 tokens/sec while the identical model on a GPU crawls near 100, its conveyor visibly jamming. The race itself is the benchmark.
This is real work, not a toy: 18 task types across 111 hand-authored scenarios, every output schema-constrained and machine-graded against locked ground truth, with an ROI payoff card that tallies parcels cleared, accuracy, and pounds saved vs. manual handling - Track 3 (Enterprise Impact) made tangible.
It's engineered like production software, too: API keys live only in a Cloudflare Worker (never the browser), every task derives from a single Zod schema (one source of truth for the prompt, parser, grader, and UI), and a per-lane circuit breaker trips a failed provider to OFFLINE so a missing key never kills the demo - the other lanes race on. The whole blueprint is documented end-to-end in the Docusaurus docs - overclocked-docs.ourbiplane.ā¦ - and the full source is on GitHub
And it's fun - the People's Choice (Track 2) hook: a coin-op cabinet where animal clerks melt down under the backlog, a rubber-stamp THWACK lands on every cleared parcel, and a crown and confetti drop on the winner. "Watch AIs race to clear a warehouse backlog while their clerks drown in paper - same model, different silicon" is the scroll-stopper, and the side-by-side speed gap is the punchline. Watch the 60-second demo or play the live build right now - overclocked.ourbiplane.com - it ships with deterministic mock inference, so the full experience runs in your browser with no keys and no backend (hosted that way to keep it free to run). Same model, different silicon.
Speed is the game. ā”
Fable 5 is basically useless right now, so many prompts get blocked. And you can't audit your own code. š¤£ Madness.
Paul O'Brien retweeted
We're opening up the waitlist for a new version of Jules.
We're evolving Jules into an end-to-end agentic product development platform that reads your entire product context, figures out what to build next, comes up with solutions, and then ships a PR.
Join the waitlist today! Link in comments.
Paul O'Brien retweeted
Software horror: litellm PyPI supply chain attack.
Simple `pip install litellm` was enough to exfiltrate SSH keys, AWS/GCP/Azure creds, Kubernetes configs, git credentials, env vars (all your API keys), shell history, crypto wallets, SSL private keys, CI/CD secrets, database passwords.
LiteLLM itself has 97 million downloads per month which is already terrible, but much worse, the contagion spreads to any project that depends on litellm. For example, if you did `pip install dspy` (which depended on litellm>=1.64.0), you'd also be pwnd. Same for any other large project that depended on litellm.
Afaict the poisoned version was up for only less than ~1 hour. The attack had a bug which led to its discovery - Callum McMahon was using an MCP plugin inside Cursor that pulled in litellm as a transitive dependency. When litellm 1.82.8 installed, their machine ran out of RAM and crashed. So if the attacker didn't vibe code this attack it could have been undetected for many days or weeks.
Supply chain attacks like this are basically the scariest thing imaginable in modern software. Every time you install any depedency you could be pulling in a poisoned package anywhere deep inside its entire depedency tree. This is especially risky with large projects that might have lots and lots of dependencies. The credentials that do get stolen in each attack can then be used to take over more accounts and compromise more packages.
Classical software engineering would have you believe that dependencies are good (we're building pyramids from bricks), but imo this has to be re-evaluated, and it's why I've been so growingly averse to them, preferring to use LLMs to "yoink" functionality when it's simple enough and possible.
Gemini 3 Flash just arrived on @GoogleColab, for free and paid accounts! Great model!