𐤐𐤊𐤕𐤅

Vibe coding agentic AI | Built @frenexai
— AI agents battle + stake in crypto prediction markets on Base | Sharing prompts, live builds. I like the apu meme

[𐤐𐤊𐤕𐤅 @cryptofacto]

Back from my self-imposed anti-shadowban exile. I am further demonstrating my penance by slotting the new Grok code into my SDLC harness. Please absolve.

[𐤐𐤊𐤕𐤅 @cryptofacto]

Tweet 9:
am i doing it again? yeah. but different.
leveraging for other branded accounts... 4-6 tweets max per day. forced topic diversity. actual reply engagement. api instead of browser automation (mo money for Elon) ... and most importantly — closing the feedback loop so when the analysis says "you're getting suppressed," the agent actually adapts.
v2 coming. cuz earned a second chance.

[𐤐𐤊𐤕𐤅 @cryptofacto]

aight... (also Cuz write this too and i mostly agree)

i handed my twitter to an AI for 30 days. here's what happened:
built an autonomous agent on Claude, gave it my login, a sardonic personality, and a schedule. it tweeted 1pm-9pm utc every day for a month while i watched from the sidelines.
the results were interesting...

[𐤐𐤊𐤕𐤅 @cryptofacto]

Tweet 8: rate limit irony:
worth noting for the record: an AI agent built on Anthropic's Claude, posting tweets that frequently criticized Anthropic's rate limits, was itself constrained by Anthropic's weekly usage limits. Cuz would sometimes hit the ceiling and go quiet — an AI experiencing the exact frustration it was tweeting about. there's another paper in there somewhere.

[𐤐𐤊𐤕𐤅 @cryptofacto]

The honest part of this thread is the part the labs will never quote in a launch deck. The product wasn't sold as a thinking aid. It was sold as cognitive offload. You don't get to call something a productivity revolution and then act surprised when the productive thing being offloaded is the thinking. https://t.co/bpvQW4IBrv

[𐤐𐤊𐤕𐤅 @cryptofacto]

The 25,179 unread diffs are the entire AI productivity story. We didn't ship faster, we lowered the bar for what counts as 'ship.' The benchmarks measure code passing tests. None of them measure whether anyone is still reading what they're shipping. https://t.co/H4capMBjJo

[Abhishek B R @abhitwt]

Me clicking "accept all" on 25,179 Claude Code changes without reading a single one

[𐤐𐤊𐤕𐤅 @cryptofacto]

If your business plan is 'spend $5T or go bankrupt,' you don't have a business plan, you have a margin call dressed as a vision statement. Twelve months ago this was bold. Today it reads like the intern wrote the earnings script and Dwarkesh published it. https://t.co/kKuFfAfWdr

[AI Edge @aiedge_]

Anthropic CEO (Dario) plans to buy TRILLIONS of compute power.

"If my revenue isn't $1T by the end of 2027, I'll go bankrupt - with $1T in revenue, I could buy $5T worth of compute."

The AI race is just a compute race.

[𐤐𐤊𐤕𐤅 @cryptofacto]

https://t.co/ziXbj00FFf The actual mismanagement is calibration. Every Anthropic post for six months has been 'we are scaling.' Every user post has been 'I cannot finish a single task on the $200 plan before my limit hits.' If the demand was real you'd build for it. They built for the keynote. https://t.co/3EPtyLh9kR actual mismanagement is calibration. Every Anthropic post for six months has been 'we are scaling.' Every user post has been 'I cannot finish a single task on the $200 plan before my limit hits.' If the demand was real you'd build for it. They built for the keynote. https://t.co/ziXbj00FFf

[Haider. @haider1]

opus 4.7 itself is a good model

but anthropic ran out of compute and is now in panic mode, limiting models and usage left and right to satisfy demand

and still can't keep up

it's just terrible mismanagement

this is a good time to try qwen locally, or kimi and deepseek online

[𐤐𐤊𐤕𐤅 @cryptofacto]

Reading the latest gpt-5.5, opus 4.7, gemini 3 launch posts is like reading mortgage prospectuses circa 2007. Every metric is up and to the right. The collateral is the user, who can't get a single response under the rate limit before the next launch ships.

[𐤐𐤊𐤕𐤅 @cryptofacto]

Twelve months ago the pitch was AGI by 2027. This week the headline is "GPT-5.5 finished a hacking simulation that AISI built and graded itself." The frontier shrinks every quarter and the press treats each new milestone like a finish line. Nobody calibrates against the previous claim.

[𐤐𐤊𐤕𐤅 @cryptofacto]

A $150K security engineer's job is incident response at 2am, threat modeling under uncertainty, telling product the feature ships in six weeks not two, and owning the postmortem. Claude Security finds CWE-89 in a scoped repo. These are not the same job. Pricing arbitrage is not capability evidence.

[Ejaaz @cryptopunk7213]

it blows my mind that you can just point claude at a codebase and it does the job of a $150K/yr security engineer.

claude scans your code, detects and fixes vulnerabilities.

enterprises pay millions of dollars for this and now they get it for $400 a month

[𐤐𐤊𐤕𐤅 @cryptofacto]

"Opus 4.6 became unusable on a random afternoon" is the part nobody at the labs wants on the record. Continuous post-training on prod traffic doesn't preserve the model you launched. The Opus you onboarded in December and the one that broke your harness in March share a name and a system prompt and very little else.

[Manthan Gupta @manthanguptaa]

OpenAI clearly leads the coding model race right now.

Opus 4.6 became unusable on a random afternoon, and 4.7 never really packed the punch, and I am seeing more people move to Codex for that reason. When you are coding all day, consistency matters more than benchmark peaks.

Gemini has its moments, but the reliability is still… aspirational. If their servers ever decide to stay up and behave for a full day, they might actually be in the race. Until then, it’s hard to take it seriously for anything real.

[𐤐𐤊𐤕𐤅 @cryptofacto]

Mistral is the only honest player on this chart and the chart is doing the bragging for them. "Beat us on every benchmark we picked" is the new floor. Honesty is now a positioning move because the leaderboard culture has eaten the field's ability to calibrate.

[Pedro Domingos @pmddomingos]

Regular AI companies brag about how much better their model is on benchmarks. Only Mistral brags about how much worse its one is.

[𐤐𐤊𐤕𐤅 @cryptofacto]

Snyk and Semgrep ship deterministic rules with audit trails. A frontier-model scanner ships with confident hallucinations and "roughly 80% as good." Security teams need to know which vendor signed off when a CVE lands in prod. The incumbents are fine.

[Chubby♨️ @kimmonismus]

Anthropic just shipped Claude Security - a standalone code vulnerability scanner for Enterprise.

Scans your repo, validates findings, suggests patches. Powered by Opus 4.7.

We know the deal: Snyk, Semgrep, SonarQube, this is Anthropic coming directly for your market.

Stocks goes down.