CLRT retweeted
Microsoft's email went down for a day last week.
Their report says why: they mostly tested on weaker machines.
The bug hit their biggest, newest servers hardest. It had been hiding there since July.
Your tests are only as good as the machines you run them on.
CLRT retweeted
An invented sentence in a transcript can look exactly like something the speaker said. In a Whisper study, about 1 in 100 transcripts contained inventions. The dangerous task is the one nobody thinks needs checking anymore.
CLRT retweeted
Did Google just double the price of cheap AI? 馃 Gemini 4 Hydrogen costs $1.50 per million input tokens, twice Gemini 3.8 Flash today, yet beats Argon on the Artificial Analysis index. For operators who budget AI by the token, not the finished job.
CLRT retweeted
You can name your new hire's manager and last working day. Can you do the same for the AI agent connected to your systems? The permissions are only the first line of its employment file.
CLRT retweeted
Is Google's strongest AI model really this cheap? 馃捀 Gemini 4 Carbon costs $2 in and $10 out per million tokens until 30 November, then it doubles to $4 and $20, per Google's pricing page. For founders and operators setting next year's AI budget.
CLRT retweeted
GPT-4o's 128,000-token window held up to about 8,000 in one test. The brochure measures what fits. Your contract review needs a different test: can it find the clause you need when the question uses different words?
CLRT retweeted
Is the best AI model now the cheaper choice? 馃Л In Anthropic's own tests, Claude Fable 5.2 nearly matched Opus 5.5's top FrontierCode score for roughly half the cost per task. For founders and operators deciding what to build, and at what effort.
CLRT retweeted
The UAE Cabinet approved 32 AI advisors and published its numbers alongside them. Abu Dhabi's TAMM says AI resolves 95% of its requests. The interesting question is what counts as resolved, because the same word can mean 14%.
CLRT retweeted
AI saved the time. A study of about 25,000 workers found no change in earnings or hours. Where did the hour go?
CLRT retweeted
85% of your AI allowance left. Reset at midnight. Whatever you invent to use it up, someone has to review tomorrow, and that person's calendar changes the calculation.
CLRT retweeted
A six-month-old runbook makes you suspicious. One updated five minutes ago gets followed. AI can fix the timestamp without fixing the truth. What would make that page safe to act on?
CLRT retweeted
Klarna says its AI did the work of 700 support agents. Then its CEO promised customers a human if they wanted one. Those decisions look contradictory until you split a support queue by what the customer is buying.
CLRT retweeted
The revenue number that lags a week. The CRM stage that means whatever the last salesperson wanted. The "temporary" spreadsheet that's been the real ledger for six years. Your team routes around all of it. Your AI agent won't.
CLRT retweeted
You finally spend your week doing what you're best at. Now comes the harder question: which part should still be yours?
CLRT retweeted
Half of the $13.8 billion US companies spent on generative AI in 2024 stayed in technical functions. Was that where the best work was, or where buying software was easiest? Put your own budget beside the work it could be doing.
CLRT retweeted
If an AI agent on your systems causes a loss, which policy pays? Most companies would say "our existing cover". Since January 2026, insurers have been able to write generative AI out of it.
CLRT retweeted
Microsoft's strategy and transformation chief: "Adding agents to a broken process still leaves a broken process."
Kathleen Hogan wrote that Microsoft first treated AI like a traditional technology rollout. What they learned: "access and usage do not equal transformation." An adoption dashboard counts logins. It can't tell you whether a single process changed.
CLRT retweeted
Everyone repeats "95% of AI pilots fail". The study behind it measured something else, and it doesn't blame the models. The choice that sank most pilots was made before anyone wrote a prompt.
CLRT retweeted
If AI got this wrong, would you notice? The wrong answer can look just as convincing as the right one.


















