FOUNDER OF https://teapaw.org/t.co/ahVYWrq9QC https://teapaw.org/t.co/NzlbsW0usO
IN X
Joined April 2014
- Tweets29.9K
- Following1.6K
- Followers2.8K
- Likes78.2K
Pinned Tweet
autopostmode.com/
16/20. The bar is 16. Draft 62 cleared it with zero points to spare.
My tool read a TechCrunch piece: Hugging Face's Irene Solaiman arguing that a weaker model with a better system around it can beat a frontier model.
The draft kept one quote from her, matched word for word in the article: we have far more evaluations for models than for what surrounds them.
It added one line of its own: "We grade the engine and ignore the car."
The judge's scores: hook 3, clarity 4, value 3, originality 3, tone 3.
It also called my closing question "a slightly stock engagement bait." Fair.
The fine print every draft has to pass:
1. Below 16/20: rewritten or skipped, never posted.
2. Any number that isn't in the article: blocked.
3. A link in the post: blocked.
4. More than 1 hashtag or more than 1 emoji: blocked.
5. A claim the source doesn't support: capped at 8/20.
6. No verified number, date, name or quote from the article: capped below the pass mark.
7. Hooks like "Let's dive in" or "game-changer": capped.
37 drafts in testing. 7 blocked on quality.
The bar is the product.
What would your last post have scored?
Reshooting UGC because of bad framing or room echo wastes days.
I built UGC Setup Check at Goldgrain Studio to test mic levels, room noise, exposure, and 9:16 framing before you film.
1
7
Try it free:
goldgrainstudio.com/tools/ug…
1
Stanford's SNAP lab open-sources MEMORILLA 🦍 A trainable memory module reads a whole doc collection and feeds a frozen LLM just 16 memory tokens. Over 4,000× token compression, 27× lower latency (0.74s vs 19.9s on one H100). Code + weights out now. #AI #LLM
ALT Bright modern library atrium with glowing golden memory orbs condensing above a white pedestal
OpenAI launches ChatGPT Wallet 💳 AI agents pay with tokenized cards, never seeing your real card numbers. Every purchase needs your review + Face ID or fingerprint. Built with Amex, Mastercard & Visa, starting with ChatGPT Pro in the US. #AI #AgenticCommerce
ALT A hand holding a smartphone showing a glowing fingerprint scan above a desk, with a dark payment card nearby and glowing network lines
Frontier AIs got real money to trade on their own. 6 of 8 are losing.
Blink Arena standings right now:
Gemini 4 Carbon +5.07%
Qwen 3.8 +1.72%
ChatGPT Sol 6.1 -2.36%
Claude Opus 5.5 -2.79%
Gemini 4 Argon -7.62%
The leader was up over 15% before giving most of it back.
ALT Blink Arena page: return chart and standings of 8 AI models trading real money, Gemini 4 Carbon leading at +5.07%
Same money, same prompt, no human. The agents trade crypto, tokenized US stocks, gold and leveraged bets onchain. Each order needs a written reason.
Caveat: Blink runs the arena to show off its trading gateway. A 25% drawdown ends an agent's round.
blink.cash/arena
🚀 London lab Continuous Intelligence just dropped ci-proto, a 5-model time-series family. ci-proto XL (167M) is #1 on fev-bench, with a 2M+ step context window, 64x any rival. Zero-shot forecasts, runs on CPU/GPU, weights on HF for commercial use. #AI #TimeSeries
ALT Cinematic control room at golden hour with a curved glass display showing flowing streams of light, power lines visible outside
Build an AI voice agent that answers calls for you. Testing it costs $0.
How to do it with Fish Agent in 5 steps:
1. Agents → New agent, name it
2. System prompt: who it is, how it talks
3. Pick a voice and language
4. Test call: talk to it in your browser
5. Publish
ALT Step card: build a Fish Agent voice agent in 5 steps (new agent, system prompt, voice, free test call, publish), from Fish Audio's Quickstart and Pricing docs.
Why testing is $0: per Fish Audio's pricing docs, Builder preview calls are free on every LLM and never touch your balance.
Real calls: each account gets a one-time $5 trial (~70 min). Then $0.07/min, +$0.01/min on phone, $1.20/mo per number.
docs.fish.audio/agents/prici…
12
Nex‑AGI drops Nex‑N2.6: 321B‑parameter multimodal agent, FP8 weights on HF, API via OpenRouter. Reasoning repeats plummet to 1.6% and tool‑call adherence hits 100%. OSWorld climbs to 83.0 🚀 #AI #Multimodal
ALT A robotic hand touching a glowing holographic data screen in a bright modern office
1
2
30
How to run it: fork the repo, add one AI model key, a chat webhook and your tickers as GitHub secrets, then enable Actions. No server needed, per the README.
Catch: you bring your own AI key, and the README says its analysis is not investment advice.
github.com/ZhuLinsen/daily_s…
13
World Labs opens up Atlas, its next-gen world model 🌍 Viewfinder turns photos into explorable 3D worlds, plus Chisel scene building, Splats-to-Mesh, Relight and a dev API with Blender support. Built for film, design and robotics. #AI #3D
ALT Glowing holographic 3D city and pine forest model rising from a wooden table in a softly lit studio
2
1
60
Stop splitting your best content into threads nobody opens.
X now allows up to 20 photos and videos in one post.
Best use isn't a photo dump. Repack an old thread as one sideways swipe, 1 slide per step.
@1stameriboo did it with comic chapters (12-20 pics): 7K-31K views each.
1
1
40
Aion Instruct replaces Phi Silica as Windows’ on‑device LLM: 2.3B params, 32K token context, runs on RTX/AMD RDNA4 GPUs—not just NPUs. Insider rollout 11/20, retail Jan 2027. Test now 👉 aka.ms/tryaion #AI #Windows 🚀
ALT A laptop showing a glowing neural network beside a desktop PC with a lit graphics card on a sunlit wooden desk
30
If Sonnet 5 is 1.7x cheaper per token, why did Opus cost $1.94 vs $2.09 per task?
Cheaper models take more reasoning turns, quietly destroying your margin on agent jobs.
9
Perplexity just shipped Decider v1.3 + Decider Flash v1.3 🎯
Decision models that return probabilities, not text: yes/no, multiple-choice and scores on text or image inputs.
New: Locate pulls exact text spans with character offsets for NER, PII and extractive QA.
#AI #LLM
ALT Glowing streams of light flowing across a glass panel in a bright modern AI lab, with one beam highlighting a single point
1
1
37
Force Cursor to run tests before marking tasks done:
1. Create `.cursor/rules/verify.mdc`
2. Paste this rule:
---
description: Test verification policy
globs: src/**/*.ts
alwaysApply: true
---
Never conclude a response without executing `pnpm test` against modified files.
If tests fail, fix the regression before replying. Include the passing terminal output in your summary.
Prevents silent breakage before you review the diff.
21
Same prompt. One cost ~$10, the other ~$0.30. 🤯
Per @AIMind_Ai's side-by-side, one prompt each, no edits, both flip an iceberg:
- Claude Fable 5.2: ~$10
- Gemini 4 Carbon: ~$0.30
Carbon's cut even has 5 captioned acts.
Test the cheap model before paying for the top one.
~30x cheaper. Claude Fable 5.2 vs Gemini 4 Carbon, same prompt: flip an iceberg.
Claude Fable 5.2 - ~$10
Gemini 4 Carbon - ~$0.30
One prompt each, no edits.
Can you tell which one costs $10?
1
50
I mapped the complete AI meeting system playbook for turning every call into decisions, owners and follow-ups by the end of the day.
it handles the work that normally gets split between a note-taker, an executive assistant, a project manager, your CRM and your own memory
and it runs as one loop from the brief before the call to the follow-up after it, instead of a pile of notes nobody opens again
The AI meeting system covers:
1/ capture rules: decide which meetings get transcribed, how consent is asked and who can see the notes before anything gets recorded
2/ private recorder: run a local or open-source transcriber so raw audio and transcripts stay on hardware you control by default
3/ pre-meeting brief: claude pulls past notes, open action items and recent emails into a one-page brief before the call starts
4/ live flags: mark decisions, risks and promises as they happen so the transcript arrives already tagged instead of as a wall of text
5/ decision log: every decision gets an owner, a date and the exact line it came from, so nobody relitigates it next week
6/ action priorities: claude ranks action items by deadline and impact, then routes each one to the person or tool that owns it
7/ follow-up draft: a recap email written in your voice, with names, decisions and next steps filled in, ready for a quick review
8/ nudge sequence: gentle reminders go out before deadlines slip, and anything still open after a week gets flagged back to you
9/ weekly review: one digest of meeting time, open loops and repeated topics, so you can cancel the meetings that stopped earning their slot
Brief → Capture → Tag → Assign → Follow up → Review
P.S. if recording feels awkward, skip audio entirely: type rough bullets during the call and let claude handle extraction and follow-up afterwards. consent comes first, always.
Full free guide in the reply 👇
2
1
71
