This 👇… 💥!

using agents to write software feels like opening a present every time they ship a new feature. it feels like xmas multiple times per day, every day.

this is gold

Happy hour at #CVPR2026 tomorrow 🍸 Come meet @a16z infra a dozen of the best generative media & computer vision teams in one place. I can promise great food, drinks, and company! RSVP👇

This was a delightful read and very timely for #CVPR !

Going to CVPR in Denver this week? Our team at @a16z and several of our portfolio companies are hosting events! Come hang with us - links 👇

If you put n dots on a sheet of paper, how many pairs of dots can be exactly one unit apart? A question that has stumped mathematicians for decades. Solved in <32 hours and $1,000 in tokens.

The new turing test

🤣🤭… I love the internet

Props to the LinkedIn rep who sold this ad. “Ya trust me we’ll get more people to use the post office… know some VCs who are clamoring for Ground Advantage.”

Pay attention to Codex. I’ve replaced Claude Cowork entirely. Note that I’m referring to Codex, the desktop app/agent harness and not the models. A few observations: 1/ It’s great for long running tasks that involve a ton of context. With other models, I notice output degradation as the number of output tokens grows. Not the case with Codex. 2/ I no longer use ChatGPT desktop app. I instead use Codex for everything. I dumped ChatGPT for Claude a few months ago bc I couldn’t get over the style issues. Even with personalization, it always sounded like AI. Codex solves this, and now it’s my go to. 3/ There are so many good commands and features triggered by “$” and “/“ — you just gotta start exploring. For example, the /pet turns on a cute little codex avatar that lives on your desktop and updates you on the state of your agents. 4/ Codex is so good for generating and editing spreadsheets with complex models. Models that would take me days to create now take only a few mins. It’s bananas.

As of today, the Federal government’s Intel position is worth ~$43 billion, representing a 385% return in ~8 months. Acquired: Aug 2025, 433.3M shares at $20.47 = $8.9B (9.9% stake) Funded by: $5.7B in unpaid CHIPS Act grants $3.2B from the Secure Enclave program (converted to equity rather than disbursed as grants) Current price: ~$99.62 (May 2, 2026) Current value: 433.3M × $99.62 ≈ $43.2B Unrealized gain: ~$34.3B or roughly a 385% return One big caveat is that the position is illiquid in practice, since selling it would move the market. Also worth mentioning: the government received a five-year warrant for up to 240.5 million additional Intel shares at a $20/share strike. The warrant becomes exercisable only if Intel no longer directly or indirectly owns at least 51% of its foundry business. With $INTC near $100, the warrant is deep in the money on an intrinsic-value basis of roughly $19B before any probability or timing discount. But it is not currently exercisable; it would only be triggered if Intel falls below the 51% foundry-ownership threshold during the five-year warrant term (late Aug 2030).

This 👇 = 🏆

The US sentiment compared to that of China is striking 👀

Our team @a16z infra is hosting a happy hour for women in AI ✨ This event is designed for researchers, engineers, and MTS to connect and chat all things emerging research. It's next Wednesday in SF - just a few spots left and RSVPs are required, see link 👇

Stop bellyaching over the incremental costs of AI features. The smartest buyers realize that making a $200k/yr employee even 5% more productive is actually worth more than $10k/yr bc it’s not zero sum. 5% (or whatever %) saved on one task translates to effort reallocated elsewhere. We’ve got to stop anchoring on software prices of yesteryear. Machine intelligence deserves the value it creates.

Unsure why everyone keeps saying “heated resin” to me today. It’s a regular Sunday best I can tell.

🚨 New: Excited to announce we @a16z are leading @deeptuneai’s Series A. RL environments are becoming both the bottleneck and the unlock for training the next generation of frontier AI models. The shift is clear: from static datasets to dynamic, engineered environments where models learn to write code, perform knowledge-work tasks, and fully control computers end-to-end. Deeptune is building this layer from the ground up and is already driving progress with top AI labs. Learn more about Deeptune and @timlupo: