Deep Thrill

GPT 5.5 turned out a steaming pile overnight and wow is anyone actually good at this yet? This is starting to feel like programming again, that feeling it’s impossibly hard and painful

**One almond ≈ 1 gallon (3.8 liters) of water** (standard cited figure from almond industry data; full footprint studies put it around 3–12 liters depending on inclusion of rain/byproducts). **Average ChatGPT query ≈ 0.000085 gallons (~0.32 ml)** per Sam Altman's 2025 figures (energy ~0.34 Wh/query + data center cooling). **Ratio: ~11,800 ChatGPT queries** per almond (3.785 L / 0.000322 L). The 15k ballpark in the quote is reasonable with slight variations in almond stats or efficiency. Real numbers depend on model, query length, location, and cooling tech—some independent studies estimate higher per-query use (a few ml to tens of ml), which drops the ratio.

I ate one almond the other day. Then I felt bad about the water usage, so I thought to myself - well I just won’t make the next 15,000 AI queries I was planning on making

The measure of code quality is the amount of time it takes to add features and squash bugs. Higher quality foundations = More rapidly improve the final product

AGI solving software means we can do the most extreme things with hardware Massive nuclear reactors, space travel and robots are just the start

Do you believe humanity will ever reach a point where we fully understand the universe? ✍️

one of the most interesting things about ai products today is that almost none of them are *live*. there’s nothing running continuously, reacting to context as it changes.. maybe a scheduled digest here or a timer there, but that’s just pull dressed up as push. everything is fundamentally a vending machine where you walk up, ask, get an answer, & then leave. getting this right is obviously tricky & the business model behind must fit to justify the burn but this is where really interesting application layer problems live rn.

software engineers before vs after AI agents

My coding workflow is currently: 1. Spend about ten minutes writing a plan for a new feature in a markdown file as bullet points. 2. Spend about an hour with a bunch of back forths between Opus 4.6 and Codex 5.4 xhigh, asking each one to improve the markdown plan file. 3. Read the final markdown plan, which ends up as a bunch of bullet points, includes implementation details, and is usually between 200-500 lines long. 4. Ask either Claude Code or Codex to implement it. 5. Ask the other one to review the implementation. 6. Run my /deslop command on the implementation (see my pinned tweet). 7. Deploy, test, and ask an agent to fix any bugs.

I cannot imagine NOT using AI to code anymore. That cat is out of the bag. Could I go back if it disappeared tomorrow? Definitely. But it exists, and refusing to use it at this point is pure self-sabotage. You're actively handicapping yourself while everyone else ships faster.

One of the most powerful use case for /goal in Codex with 5.5 xhigh: performance optimization In my game I display >100k moving items and 10k moving animated zombies. And I couldn't understand why I was at 120fps average but with big dips to 20fps at random time. I was going crazy. So I launched Codex and it spent THE WHOLE DAY on this "/goal run this 5s gameplay save: profile and iterate until you strictly reach 0 frame above 8.3 ms" Now my game is BUTTERY smooth, average 200fps with 0.01% at 120fps. Impossible optimization is now possible