Building AI products and figuring out what actually works.
I share what I’m learning about:
• AI products
• Agent systems
• Growth & execution
If you're building, shipping, or experimenting with AI, follow along.
@sysoft_cn@thsottiaux Nope. Same as always: 80% coding, 20% product work. The only new thing was using the free API credits for a data analysis demo, which cost less than $20 total.
@codewithantonio@instinct@bot Same here.
From what i’ve read, it’s the deep access to email/calendar plus the freedom to act without asking at every step. and it knows when not to text, which matters just as much
@EMarkosov Self-hosting the harness and session history helps with infrastructure control. But if the blocker is prompts leaving your network, where model inference runs still matters. Those are separate requirements.
Google's Gemini Decisions API:
Decision 1 leads 5 of 7 shared tasks in Google's own chart, not all.
Promising, but the real test is fewer wrong tool calls in production—not just another benchmark win.
https://t.co/PZIDXUVdNF
@elonmusk@bot@elonmusk@Bot My Grok Bot computer has been unusable for ~2 weeks after a storage issue. Disk Saver, Reset and Update all fail, and the cleanup tool itself is broken, so I can't fix it myself. Ticket T-I86600 still has no recovery plan. Could the team escalate it? 🙏
Cursor Origin, Entire, Tangled. A new batch of "GitHubs" built around agents.
Now GitHub is testing Git storage for thousands of concurrent agents.
The latest entrant building the next GitHub: GitHub.
https://t.co/B8ZN5Jq9QT
Anthropic published how Claude agents build memory and why dreaming matters.
Agents write notes while working. Offline, dreaming rechecks them against transcripts, fixes stale ones, and builds a wiki.
I'm treating memory cleanup as part of the harness.
https://t.co/dc8QIfZ9xk
An agent going idle doesn't mean the job is done. Its shell command may still be running.
I want the final command result before the agent reports success—not just a "completed" notification.
Three Images 2.6 GIF ideas worth trying:
A logo made of glowing text: https://t.co/jX9Lt5byM2
Reaction GIFs: https://t.co/HffGjoTzSh
Looping character animations: https://t.co/ulv5lhQps1
I'd start with loading states and small UI animations.
Another tip: Images 2.6 can now make GIFs directly. The image shows the same character as a 12-frame looping GIF at 𝗹𝗼𝘄, 𝗵𝗶𝗴𝗵 𝗮𝗻𝗱 𝗺𝗮𝘅.
Low has the biggest motion, but the face and lines are clearly blurry. Max has the sharpest hat and eyes. On a phone screen, you can barely tell high from max.
Cost per GIF through the API: low about $0.055, high $0.12, max $0.47, and max takes two minutes. If you use Codex, 𝗶𝘁'𝘀 𝗳𝗿𝗲𝗲, though you can't tell yet which level it uses.
Attach your character image and paste this into Codex 👇
"Turn this character image into a seamless looping GIF, 12 frames at 100 ms per frame. Action: the character spins the staff in their hand, releases a small burst of sparkles, then returns to the starting pose. Keep the design, colors, proportions and line art the same as the original in every frame. Full body in frame, fixed camera, plain pure white background, no text."