Joined December 2017
- Tweets 1,891
- Following 190
- Followers 64
- Likes 1,591
70 Photos and videos
Pinned Tweet
May 25
Most AI accounts either sell hype or talk over your head. I do neither. I test new tools, break things, and tell you what actually works. Not a developer. I figure things out with AI, not around it.
1
3
597
10h
Minimax M3 with 1M context running open weights on consumer hardware is the sleeper I didn't see coming. Throwing full codebases or long research docs at it locally, no API calls, changes how you experiment. Quantization on Mac feels like the practical next step most people will actually run.
Anyone testing it quantized yet? Real-world speed on Apple silicon?
31
13h
Aster claims thousands of parallel AI agents helped them hit a ProteinGym world record in 30 minutes with massive research speedups. The YC batch is full of these "AI workforce" plays. Parallel agents sound impressive but the coordination and error-correction overhead is the real hidden cost nobody talks about. A benchmark win may not translate to messy real-world research. For actual tinkerers, starting with 5-10 well-orchestrated agents feels far more practical than chasing thousands. Is the thousands-in-parallel claim realistic or just good marketing?
25
16h
1/6 DeepMind dropped a 57-page report arguing AGI won't be a stable endpoint. It triggers fast cascades to ASI through multi-agent "hive minds." Made me rethink how I'm building agent systems.
1
28
16h
5/6 The skeptical note that keeps nagging at me: solid research, but real agent swarms still choke on coordination overhead and error correction the paper mostly skips. Theory and messy deployment are still far apart.
1
17
16h
6/6 This pokes a hole in the "AGI as endgame" story. Does the hive mind framing change how you're building agents, or does it still feel too far out?
15
Jun 13
Google's AI is reading your Gmail by default now. Emails, attachments, the bank statement your accountant sent in April. I shut it off the minute I saw it.What gets me isn't even the feature. It's that it's on unless you know to turn it off, and the setting is buried where most people will never look. That's not "helpful AI." That's a default nobody agreed to.There's already class-action talk floating around, so apparently I'm not the only one who finds this creepy. So: paranoid, or fair? Curious how many of you are turning this off too.
48
Jun 13
1/6: That "10 GitHub repos for AI automation in 2026" list is getting bookmarked everywhere. The replies are better than the post though. Reading replies first is basically my workflow now.
1
50
Jun 13
5/6: It's fine as inspiration. I'd kill for a simple test protocol though, instead of trusting star counts. Would that change how you pick tools?
1
39
Jun 13
6/6: Which ones have you actually kept using? Or did they end up in the "looked cool, didn't stick" pile like most do for me?
28
Jun 13
NVIDIA published a benchmark showing Blackwell beats Hopper 20x on agent workloads. NVIDIA ran it. The unit: per megawatt. Not per dollar, not per token. Real infra decisions aren't made in watts. When the vendor runs the test that validates what it sells, what should we call that exactly?
45
Jun 12
1/4 Two numbers from this spring that belong in the same sentence: model prices keep dropping toward "basically free," and a single next-gen NVIDIA Rubin rack now reportedly runs about $7.8M, up from roughly $4M for the last generation (Morgan Stanley).
1
1
126
Jun 12
3/4 So the labs are selling tokens cheaper while the thing producing those tokens costs nearly double. That gap is being filled by investor money, not by the economics actually working.
1
1
82
Jun 12
4/4 "AI keeps getting cheaper" is true at the price you pay and false at the cost to produce. One of those numbers is subsidized. Worth remembering which, before you build a business on prices that assume the subsidy lasts.
1
69