Building systems, decoding protocols, and auditing every line of AI-generated code. Reverse engineering · Codex · Ollama - Ida
Joined July 2019
- Tweets164
- Following77
- Followers8
- Likes467
Pinned Tweet
MF is a management/tycoon simulation game about founding and scaling an AI company. You start in a garage in 2016 with a tiny team and grow all the way to a frontier AGI lab, deciding strategy, hiring, research, data sources, compute infrastructure, more.
github.com/eduhcastro/ModelF…
I think this setup would work well for a lot of people.
I use Codex GPT ($100 plan) as more of a real assistant for day-to-day work — controlling my computer, investigating bugs, handling deployments, working with containers, doing research, thinking through infrastructure strategies, and a lot more.
I use Claude ($200 plan) mainly for programming, both backend and frontend.
I’ve finally found a good balance for my daily workflow.
Claude doesn’t really control your computer the same way Codex does, while Codex isn’t as strong at programming as Claude.
I’m thinking about subscribing to Claude Max 5x now. Is it worth it? How are the usage limits? I have OpenAI’s Pro 5x plan. Are the usage limits the same, or are they different?
The global reset happened today, and the compute that @thsottiaux says is available is already struggling. (SOL)
In today’s livestream for the “Building a Company in 3 Days with Grok Bot” challenge, the GrokBot team was using Cursor with the Fable 5.1 model to build the website/project’s source code.
From the looks of it, Grok 4.7 may still be quite a while away, especially since even the company’s own team doesn’t seem to be using it yet.
I’d like to know what happened here. I don’t have any overdue or unpaid plans.
@spaceship @ZohoBooks
For those using Codex on the 5x plan, which model is performing well while also consuming usage at a reasonable rate?
I’m mainly doing full backend work — I don’t really care about frontend.
I’m on the 5x plan, using GPT-5.6 Sol High.
So far, a single prompt has already used 5% of my limit. Is that actually normal? What’s going on?
You see one post that’s different from the content you normally get every day, you like it, and suddenly your entire timeline stops showing the kind of content you usually follow and starts showing only that new type of content.
How do I turn this off?
Nice website
model rumor site just got a massive smoothness performance boost thanks to Astra!! (click and drag that timeline)
Also Grok 4.7 this week 👀 congratulations to Grok Bot
modelrumor.com/
GPT Sol usually takes around 30 minutes to do something that Astra can finish in 8, for example. What happened to Sol?
I’ve already tried this with both GPT and Grok. According to them, they’re not allowed to copy, download, help extract, or provide anything that contains the original text of an article, because they consider it a copyright issue, even if the article is publicly available.
I know this because both of them had access to an indexed website that is no longer online. It contained a very detailed article about large-scale deployment that I wanted to read, but they refused to provide the original content. They would only offer summaries, context, or assumptions based on what they could find.
So it’s possible he was banned for doing something like that.
A cursor user with 1.5 years and a paying account got permanently banned over a personal script.
The only thing he'd asked for was a script to download articles into his own Obsidian vault.
No refund, no real explanation, just an automated message citing a Terms of Service violation with no specifics.
What actually happened:
> asked cursor to help write a script for downloading and organizing articles
> got banned within an hour, no warning first
> support ("Sam," an AI bot) refused to name the actual policy broken
Other users in the comments found the likely cause: cursor's policy flags "automated scraping scripts" as malicious code generation, even for personal, non-malicious use.
Multiple other longtime paying users reported the exact same thing happening to them around the same time, one after building a WordPress plugin, another after routine subscription upgrades got mistaken for refund abuse.
The irony landed hard in the replies: banning a user for writing a scraper, while the AI itself was trained on scraped data at a much larger scale.
Una retweeted
We are investigating reports of degraded performance for Issues and Pull Requests githubstatus.com/incidents/7…
Una retweeted
Two weeks ago, I resigned from OpenAI to join Conduit as a founding researcher, where we're training models to non-invasively read the human mind.
I've written some thoughts about what telepathy could look like by 2035 and how to get there:
