·
Finding signal on Twitter is more difficult than it used to be. We curate the best tweets on topics like AI, startups, and product development every weekday so you can focus on what matters.
PLUS: Models learn tricks never in training
LLMs can perform tasks never demonstrated in training data. Compositionality, not just memorization, drives emergent capabilities.
AI coding tools need pinned messages and scratchpads to track what you're not ready to ship yet.
With GPT 5.5 and better tooling, prompts that took an hour now run 4 to 10 hours. Yielding agents is a skill.
Hermes Agent now saves 14% on input tokens during file reads. Free speed on every operation.
Anthropic is an intelligence company, not a coding company. Coding was just the beachhead.
OpenClaw's latest release is 14.5% faster on cold turns and 52.8% smaller on install. Lighter core, sharper claws.
Dan Shipper hit 38 billion tokens and a 56 hour task on a 41 day Codex streak. Absolute unit.
Marc Andreessen worries broken sectors will eat AI productivity gains the same way they did with computers.
Tenet Industries is building F-150s for defense while primes are still building Bugattis. Mass production beats artisan craftsmanship.
Tyler Cowen thinks people will do religion solo through AI. Why ask your priest when the model knows more?
Meaningfully better AI releases are accelerating, especially from OpenAI and Anthropic. The timeline only tracks 3+ point jumps now.
Code coverage is a vanity metric. Functionality coverage with Playwright plus MCP catches bugs users actually hit.
Slop isn't a prompt problem, it's a systems problem. You need an eval loop, not better instructions.
Opus 4.8 hit 7th on SnakeBench but still can't plan 5 to 6 steps ahead. Same failure mode across all models.
Hermes Agent fixed the laggy TUI startup hang caused by MCPs. Bug fixes and reliability all next week.
Hermes ships with a pre-bundled Polymarket skill. What the hell is happening.
You can now optimize your LangChain chains with GEPA. Community PR made it happen.
Once you internalize auto-research with /goal, you unlock ridiculously fast app improvement. We're too rooted in expensive labor.
Opus 4.8 produces different output in Claude Code versus regular Claude. The latter's still better for writing.
Voice-controlled computer navigation with GPT Realtime 2.0 is very underrated. No hands required.
Trillions in enterprise value sits in companies under 5,000 people. We're decoupling value from labor.
As models get better at coding, they get worse at strategy. Interesting tradeoff emerging.
The only thing worse than having the CEO knee-deep in AI is not having them there at all.
Work trials over weekends show candidates the real intensity. Office full on Saturday? We're not joking around.
Josh Pigford built 5 products solo with AI agents using a 3-phase skill and adversarial code reviews.
Ship the best product. Use lots of AI, some AI, maybe no AI. Just be the best.
Self-driving cars create literal standoffs in the street. Competing SaaS products would never.
GPT-5.5 described someone as competitive, but mostly against confusion. That's a banger of a read.
You only get angry at customers when you don't have a lot of them. Scarcity breeds resentment.
Codex CLI's tab key is the single best feature over Claude Code. Crazy they haven't copied it yet.