> Better Caching – Cache misses are the easiest way to drive your cost up. All of our requests are cache aware, so we’re reusing a warm cache wherever possible.
We do this for you in Deep Agents - see our blog on it here: https://t.co/a1zcMR7Sqy
Example: instead of burning thinking tokens during rush hour, Claude Code moves the thinking into the output by explaining its thought process as code comments.
Simple tokenomics 101: All tokens are equally expensive, but some tokens are more expensive than others.
Right now we pay a flat rate per token, but providers unfairly play with token quality as a new form of traffic shaping during high traffic rush hours. ...
💯the primary enterprise moat is forming around who owns the state & long term execution layer.
I work with enterprise AI leaders every day, they want control over the execution layer and want to avoid vendor lock-in, particularly with model providers. At the same time they want lightning speed process encoding and agent deployments. The next major enterprise tech player will take over this layer.
Mr. President, decisions about targeting infrastructure carry massive human consequences.
No power → no water, healthcare, food distribution
No bridges → no movement, trade, emergency response
Result → society partially collapses. Millions of innocent lives are affected.
Respectfully, this impact must be at the center of any decision. 🙏@POTUS
My race against a border ban for a shot at Silicon Valley dream, and what it reveals about immigrant founders:
https://t.co/E8ySJsyCXC
On January 27, 2017, I almost didn’t make it to San Francisco. I had a visa to enter the U.S. I had an incredible career opportunity. And for a moment, it looked like neither would matter.
Today, thousands of Iranian students already in the U.S. are facing similar uncertainty. Their legal status is unclear, even as they study, build, and contribute here.Many are in STEM fields. Many are working toward building the next generation of companies. Right now, they are in limbo.
Experiences like explain something deeper about immigration, and about immigrants. It shows why immigrants, who go through so much, end up playing such an outsized role in building our economy and driving innovation.
Simple experiment: I matched Claude Code quality on a simple 1M-line repo by optimizing prompts via AutoResearch.
I used Claude’s own system prompts and 100 Claude-written PRs for training.
The result? Qwen 3.5 jumped from 4/10 to 10/10 correct solutions (though Opus still ...
feels slightly better).
Disclaimer: This is a "toy example." Likely overfitted to this specific repo, and my agent is much slower than Claude (no concurrent LLM calls)
Simple experiment: I matched Claude Code quality on a simple 1M-line repo by optimizing prompts via AutoResearch.
I used Claude’s own system prompts and 100 Claude-written PRs for training.
The result? Qwen 3.5 jumped from 4/10 to 10/10 correct solutions (though Opus still ...
Georgi on why it's still hard to get great coding agent performance from local models:
"Note that the main issues that people currently unknowingly face with local models mostly revolve around the harness and some intricacies around model chat templates and prompt construction"