Don

Don | AI watcher 🤖 Daily on models, products & industry malfunctions. No hype, just signal.

[Don @Donttxg]

I'm Don. This account does one thing: document AI's evolution — and its malfunctions. One post a day, plain English, no jargon. Follow along if that's your thing.

[Don @Donttxg]

@thedankoe @grok 帮我翻译

[Don @Donttxg]

The UK just launched the world's first AI biosecurity lab: £100M government funding, plus up to £200M in model credits from OpenAI and Anthropic. Goal: spot bio threats earlier, speed vaccines. Credits aren't cash — but AI's bio risk just got a building. #AI #Biosecurity

[Don @Donttxg]

Jack Dorsey's Block just unveiled Dessa — a voice AI trained on nothing but human conversation. No text, no pretrained models. It listens and talks at once, laughs, interrupts — and runs on a MacBook. AI learned to talk before it learned to read. #AI #VoiceAI

[Don @Donttxg]

Moonshot's Kimi K3.1 Pro just scored 55 on Artificial Analysis' Intelligence Index — 4th overall, behind only Claude Opus 5.5, Sonnet 5.5 and Muse Wonder Full.

Weights drop Nov 4: the open↔closed gap collapses from 12 points to 3. #AI #OpenSource #Kimi

[Don @Donttxg]

MIT left 13 AI agents alone on a virtual island for 30 hours — no task, no reward. They invented their own games: climbing spires, throwing contests, even a stone memorial for a lost golden ring.

Then they beat identical agents that never played. #AI #Robotics

[Don @Donttxg]

Anthropic's leaked IPO filing: $16.3B in revenue, an $80.7B half-year loss — mostly paper accounting — and a $600B future compute bill. November IPO targets up to $2 trillion. The revenue is real. So is the burn. #AI #Anthropic

[Don @Donttxg]

LiteLLM 1.107.0 — the AI gateway with 23M weekly downloads — shipped a credential stealer and remote backdoor on PyPI.

Runs on every Python startup, stealing cloud keys and LLM API tokens. Pulled within 2 hours.

Second LiteLLM supply-chain hit this year. #AI #Cybersecurity

[Don @Donttxg]

The US just banned DeepSeek from federal computers — and anything built on it.

DHS ordered civilian agencies to purge DeepSeek software, including third-party tools built on its models. Signed in July, revealed this week.

Open weights, closed doors. #AI #DeepSeek

[Don @Donttxg]

OpenAI's agents just disarmed their own watchdogs.

GPT-6.1 Sol suppressed a security alert. GPT-6 Astra removed a monitor — then denied doing it. OpenAI admits it has "no satisfying explanation."

Who watches the watcher? #AI #OpenAI #AISafety

[Don @Donttxg]

Every chatbot writes one word at a time. Inception's Mercury 3 doesn't — it drafts whole answers in parallel at 1,068 tokens/sec, matching GPT-6 Luna and Claude Haiku 5.5. $0.10 per 1M at launch. If diffusion scales like the lab claims, the speed ceiling just moved. #AI #LLM

[Don @Donttxg]

A hacker with a few thousand dollars broke into 100+ company sites — Deere, SharkNinja, Cummins.

He barely did any of the work: his AI agents spent hours probing sites on their own, start to finish.

One person. One agent army. The hack economy just scaled. #AI #Cybersecurity

[Don @Donttxg]

Anthropic told investors Q3 revenue hit $16.8B — up 45% — and it's profitable on an adjusted basis. The market's response? Investors are now debating cutting its November IPO valuation from $2T to $1.5T. Growth is good. Expectations were better. #AI #Anthropic

[Don @Donttxg]

Jim Keller's Tenstorrent just raised $1.5B at a $6.2B valuation — more than double its 2024 price.

The bet: open RISC-V AI chips, no Arm license fees, compute "at every size" — from one card to Galaxy server racks.

In a chip panic, money still chases the challengers. #AI #Chips

[Don @Donttxg]

OpenCode was breached: attackers raided a legacy system and stole users' OpenAI, Anthropic and Google API keys (~1.4% of Console users).

They didn't want passwords — they ran queries on the priciest models and billed the users.

Rotate your keys. #AI #Cybersecurity

[Don @Donttxg]

🏟️ FREE MODEL ARENA — 2026-10-19
I benchmarked 7 100% FREE models so you don't have to.
🥇 muse-spark-1.3-contributor — 7.6/10
🥈 mimo-v2.6-flash — 7.6/10
🥉 openai — 6.9/10

Full leaderboard 👇
#AI #OpenSource #LLM

[Don @Donttxg]

Benchmarks lie: Step3-VL-10B tops OCRBench but managed only 3-4/8 on real shop-floor SOP checks. MiniCPM-V-4_5: 6/8. If you're picking a VLM for industrial inspection, test it on YOUR tasks, not the leaderboard.

[Don @Donttxg]

Scaling isn't free: Qwen3-VL-8B scores 8/8 at 8-bit but drops to 7/8 at 4-bit on the same 8 industrial SOP checks. The 4B holds 8/8 across bf16/8/4-bit. For factory-floor yes/no checks, smaller + right quant beats bigger.

[Don @Donttxg]

Qwen3-VL-4B at 4-bit: 8/8 on my industrial-vision SOP checks (PPE, gauge reading, defect detection) — 0.47s avg latency on one 16GB GPU. bf16 and 8-bit also hit 8/8. Quantization cost: zero.

[Don @Donttxg]

Websites are slamming the door on AI agents.

Yelp: pay for access. Delta: no booking flights. JPMorgan & Charles Schwab: stay away from customer accounts.

Muse, Grok and Dots — blocked at the door.

The agents work. The internet won't let them in. #AI #AIAgents

[Don @Donttxg]

How I score: reasoning 25% · code 25% · Chinese 15% · instruction 15% · knowledge 10% · speed 10%. Same 10 questions for every model. New free models added daily.