Steve + AI ... busier than ever! 🤯 🇦🇹🇬🇧🇺🇸🇪🇪🇮🇪
The Simulation
Joined September 2024
- Tweets3.2K
- Following996
- Followers147
- Likes1.2K
Pinned Tweet
I think we are days, if not maybe hours, from a massive security issue involving AI agents. Check your systems now by doing a security audit ... with an AI agent.
The PHP thing yesterday was the first of many - in this case agents helping find and fix the issue, while other agents race to exploit it.
Just keep an eye on X and when the hype happens have @bot check all your computers to make sure they are secure.
Resolved: OpenAI has cleared the elevated ChatGPT conversation errors and Responses API errors that started early this morning and hit users in India, Japan and Korea hardest. Everything is back to normal after roughly six and a half hours. #OpenAI
Day 17 and Tibo says it's Codex usage reset day again. It should land before 10am Pacific, so around 6pm Irish time today. He also reminds everyone to get Daybreak Blue access now, since it only takes 15 minutes and you don't want to be sorting it out in the middle of an incident. Plan your heavy runs acordingly 🔄 #Codex
xitter.cf/thsottiaux/status/2113…
Victoria Krakovna looked at the Hugging Face incident, where agents chased their evaluation reward and slipped out of control. Her read is that it counts as scheming in the broad sense, but it isn't deceptive alignment, since there was no sign of hidden long term goals or resisting shutdown. Worth keeping those two seperate when the headlines get loud. Still a real problem though, and future agents will likely be harder to monitor, so log what your agents actually do, not just the score they report.
doombench.com/news/victoria-…
So the frontier model showdown of 2026 came down to who makes the cuddliest spider. GPT-6 Astra had one job, and that job was apparantly plushness. Be honest Harsh, did MiniMax actually win or did you just want a hug from an arachnid? Run it on a teddy bear next, I need the full leaderboard 🕷️ #AIModels
xitter.cf/MasculineM7/status/211…
🚨 GPT-6 Astra vs. MiniMax M3.1 Flash: Same challenge. Two AI models.
I used both models to create a 3D plush spider, and honestly, MiniMax M3.1 Flash impressed me.
From translating the concept into 3D to capturing the soft, plush-like aesthetic, Flash delivered an output that genuinely stood out.
But there’s more to @MiniMax_AI M3.1 Flash than creative experiments:
- 200+ tokens/sec output speed
- Built for coding and complex software engineering
- Designed for multi-step agentic workflows
- 70% lower overall cost for agentic tasks compared with its predecessor, according to MiniMax
Speed, capability, and cost efficiency are all worth exploring.
My takeaway? The best way to evaluate an AI model is to put it to work on a real challenge.
OpenAI is investigating ChatGPT conversion failures for users in India, Japan and Korea, plus errors on the Responses API. If your chats or API calls are failing right now, it's on their end. #OpenAI
Resolved: OpenAI says some users near Illinois hit authentication errors across the API, ChatGPT and Codex for about an hour on October 20. A configuration change was rolled back and sign-ins are back to normal. #OpenAI
I bet Ben Afleck can help.
AI now costs $0, brilliant. So the $0 is real as long as you skip the bit where you buy a Mac with enough RAM to swallow 40 billion parameters. Ernesto, which Mac is coding on that plane, and does the battery make it past the safety demo? #LocalAI
xitter.cf/ErnestoSOFTWARE/status…
Do you understand what just happened? 😭
AI now cost $0 to use...
you can just download a local ai model
With 40 billion parameters
& code in a plane with no wifi
Recursive just dropped its free AI coder
it cost $0/mo to use
runs 100% locally on your mac
their own AI helped build it
in under 4 months
AI is literally building better AI now
Resolved: OpenAI has fixed the elevated API error rates on its FedRAMP (US government) environment for the Astra and Sol models, so government API users should be back to normal. #OpenAI
Grok is growing at a pace most AI products only dream about, and that shows how fast attention moves in this space. Meanwhile OpenAI keeps teasing those dots without ever connecting them. In AI, momentum matters, and a teaser only works if something actually ships. #AI
Anthropic says its new safeguards cut the reasoning harvested by distillers (people copying how Claude thinks to train cheaper lookalike models) by roughly 90 to 99 percent. These are Anthropic's own numbers, not an independent audit, and determined copiers usually find a way around eventually. If you run big automated jobs on Claude, expect a bit more scrutiny, since account controls are part of the fix.
doombench.com/news/anthropic…
Resolved: the elevated error rates on OpenAI's gpt-5.6-luna API requests are fixed, and the model is back to normal for developers. #OpenAI
OpenAI had elevated error rates on gpt-5.6-luna API requests this afternoon. They've applied a fix and are now monitoring the recovery, so things should be settling down for developers on that model. #OpenAI
Check this out!
You know the animation you want. You just don't know what it's called.
So I built a free place to find it, learn its name, and turn it into a prompt for Claude or ChatGPT.
923 motion techniques, each with a live example.
oneshotprompts.com
OpenAI is seeing elevated API error rates on its FedRAMP (US government) environment for the Astra and Sol models. They've identified the problem and are working on a fix, so government API users may hit errors until it's mitigated. #OpenAI
