Brandon Farley

AMD Senior AI Director confirms Claude has been nerfed. She analyzed Claude's session logs from Janurary to March: > median thinking dropped from ~2,200 to ~600 chars > API requests went up 80x from Feb to Mar. less thinking and failed attempts meaning more retries, burning more tokens, and spending more on tokens > reads-per-edit dropped from 6.6x → 2.0x. model stops researching code before touching it. > model tried to bail out or ask "should i continue" 173 times in 17 days (0 times before March 8). > self-contradiction in reasoning ("oh wait, actually...") tripled. > conventions like CLAUDE.md get ignored because there's less thinking budget to cross-check edits > 5pm and 7pm PST are the worst hours, late night is significantly better. this means the thinking allocation is most likely GPU-load-sensitive.

If Claude code Opus tells me one more time it’s getting late at 2pm in the afternoon I’m done. They really did make it dumb and lazy for Mythos compute. And 5.4 is great but spends so many tokens trying to save face when it’s wrong. 2026 problems.

@OfficialLoganK @_andreantonelli @GoogleAI any update on this? GCP support has no idea what gemini even is and all of our production usage was nuked to tier 1.

@_andreantonelli @GoogleAI there was a change which rolled out Friday and impacted lots of people who are on startup credits, this is being rolled back as of last night, will take ~16 hours to finish the rollback and then tiers will be restored, sorry for the hassle on this! Will post mortem.

hey @googleai my gemini api project got silently downgraded from tier 3 to tier 1. no billing issues, no violations, no notification. 429s everywhere. tons of devs reporting the same since yesterday. manual tier upgrade option is gone too. what's going on?

I thought I would take the night off but then Google dropped a new embedding model and I was too interested to see what new it understood about culture. Wrote up some initial high level results and there are definitely some interesting differences. Spoiler: It knows about our slop and brain rot.

I dig this Strava trend

One area where AI matches my coding style well is its tendency to name scripts "_final" and then create 37 more numbered iterations after.

I'm done making proclamations about model character and attributes, as frontier labs keep tweaking them for better economics and user engagement. Gemini might be the most sycophantic model right now.

Anthropic, like other VC subsidized token loss leaders, is likely adding limits due to a poor biz model fit, not to address a few bad actors running at maximum capacity 24/7 as claimed. Existing limits already targeted that, and blanket restrictions are a hammer not a scalpel.

A T-shaped career in the AI era: Deep enough to direct the model. Broad enough to connect the dots it can’t see.

@pgflow_dev @supabase This is such a great demo

This demo app scrapes a website then summarizes it, extracts sentiment and keywords in parallel - all orchestrated on local @supabase + pgflow. No servers, no cron, no SQL - just TypeScript and reliable bg jobs. Told you it’d be fun 🥳 #buildinpublic #AIAgent #AIAgents

the other thing about gemini Pro 2.5 is that it understands you yet isn't just agreeable

after coding with sonnet 3.5 almost entirely, not liking 3.7 at all except for planning (even non-coding its not clear what model is good at)... Gemini Pro 2.5 is another level. Where 3.7 and even 3.5 go off track, gemini stays with you understanding your goal AND codebase.

The plan? We’ll start a sparkling wine business out of Champaign, IL

The biggest issue is they artificially restrict innovation in order to keep their product lines from getting even remotely close to each other. this is the result. Incremental updates for everything. No touch on laptops, nothing of any depth on iPad that would resemble a laptop workload. No foldables that would threaten iPhone.

@evanochsds @VicVijayakumar What’s the secret to get apple tv to stop suggesting content on apps you don’t have, when apps installed actually have the content? To me it just seems like they care more about new subscription revenue than helping you find anything.

@VicVijayakumar If you have an Apple TV and sign into each streaming services app the Apple TV app will then be populated with the content from each different streaming service. It rules.

So sick of paying for 6 streaming services & when my kids ask for a movie I have to search each one to see who has it. Someone invent a thing where I pay one price for everything. And every now and then I’ll call you to say I don’t like your new price give me the old price.

Just as life evolved from simple to complex, AI's understanding may ultimately emerge from pattern matching after all.

Language = patterns so it’s no surprise that was the first domino to fall. Biological processes like DNA are patterns. Why can't AI's understanding ultimately come from an advanced form of pattern recognition, just like human cognition?

This whole argument that LLMs only operate on probability distributions learned from data, not on any genuine comprehension of meaning or concepts. It comes down to understanding vs pattern matching, but why does it have to be either or?