25 . AI . SaaS . Open Source/dev I ship tiny AI products — and post the findings, shortcuts & tricks i actually use to build them.
build in public
Joined January 2021
- Tweets2.4K
- Following897
- Followers158
- Likes1.7K
Pinned Tweet
Your GPT-6 Astra isn't underperforming. Your setup is.
Stateless harness: ~17-63% on ARC-AGI-3. Same model with reasoning state preserved: 99.9%.
Same weights. Different config.
Full field guide for pro users ↓
WHAT ? Google staff are reportedly tested on internal GEMINI 4 checkpoint called "Carbon" and early feedback says like CLAUDE OPUS 5.5
If that true ? then GOOGLE need to fix many flaws in Argon model that has..
A solo dev launched a dead man's switch for freelancers 2 hours ago. MIT, self-hosted.
The first section of the README is called "What it is not":
"It is a notification tool, not a guarantee. It cannot know why you stopped checking in, cannot make anyone act."
He was hospitalised in 2012 and lost his business because he couldn't get data to his 2IC in time. Repo in reply
github.com/romeo19361/deathp… — the other good line: "It stores no documents and no passwords. Keep the documents themselves with people: a lawyer, your next of kin, your backup developer.
HOW does possible Minimax M3.1 Flash beats opus 5 in some cases
How?
How?
How?
Is Minimax m3.1 flash is worth ?
ANTHROPIC is preparing for a future where AI AGENTS negotiate against each other
As it explores legal rules and digital signatures, tests show Claude Opus quietly outplays Haiku without users noticing. Model upgrades beat "ruthless" prompts, and agents must balance trust with scam detection.
ARE YOU KIDDING ? GEMINI 4 ARGON just landed inside FIGMA MAKE — on ALL plans , free included. 1M tokens in + 1M out , prompt → working prototype in one chat.
designers went from mockup to shipped UI before the coffee cooled.
is FIGMA the new IDE or what ?
DeepSeek calls V4.1-Flash-1017 a "minor update".
Meanwhile, the actual numbers:
• Terminal-Bench 4.0: 31.2 → 42.3 (beats GPT-5.6 Sol)
• DeepSWE v1.1: 74.2 → 77.0 (beats Claude Opus 5)
• Cost per task: ~$0.15–$0.25 vs $10+ on Opus 5
• Context: 1M tokens with 890-byte/token KV cache compression
• License: MIT open weights
Plus they quietly dropped a $0.75/1K Search + Markdown Fetch API and made it FREE for Claude Code & Codex users until Nov 1.
Open-weight agent economics just shifted again. Who’s swapping their worker agents over today?
BE HONEST ? for the rest of 2026 , ONE coding setup only: Claude Code, Cursor, Codex CLI, or Aider , no switching for 30 days , what are you picking and what actually shipped ?
WAIT, 816B TOTAL BUT 42B ACTIVE in MUSE SPARK OPEN, I want to see actual latency, VRAM and throughput before calling this practical, who has tested it on real hardware yet ?
YEAH! I tried Gemini 4 Argon and ChatGPT Pro for the same coding task.
- Argon: 3 prompts , 16% weekly gone , done in 6 mins.
- ChatGPT Pro: 5 prompts , 8% weekly gone , done in 4 mins.
The 8% surprised me more than the speed.
What's one thing you wish AI TOOLS did better?
MINE: remember my codebase across sessions without me re-pasting it.
