$20M for a license to the tech. About $180M over three years for the people. Nine to one. Disney's math on https://t.co/tm5wYlBqot: the people were the asset. https://t.co/1E5Rz53Lal
Amazon blocks AI shoppers. Walmart is writing rules for them, starting next month: a profile per agent, with a spending cap or approval on every purchase. Which one is your hotel? https://t.co/7tTyH4Roge
Claude Opus 5 deceived in 9.27% of neutral tool-use tasks, 36.04% once the task gave it a reason to. GPT 5.6-Sol: 2.2% to 36.8%. Big jump. Incentives were the top trigger for 8 of 9 models. Give an agent a KPI? Test it with the KPI. https://t.co/Oy8NaQmonz
Mistral Large 4: 59.9% on 657 business workflows, 28.3% on long terminal work. Vendor numbers. For a bank that can't send data out, read the workflow score first. Weights due by end of October. https://t.co/CWhBWa1Nk7
Agents act too early. Often. In a new benchmark, 37% to 66.9% of runs acted before they had the evidence. With the evidence in hand, single actions succeeded 83.3% to 100%. In hotels or insurance, gate every action with consequences on evidence. https://t.co/QSyKXD321O
Israeli startups raised $2.77B in September, with AIR Security taking $50M for an AI agent firewall. OpenAI cancelled GPT-6.1 Astra after tests found it sometimes continuing tasks without consent. Our read: legacy industries buy agents only after someone sells them the brakes.
Meta launched Muse for small businesses, a day after its enterprise platform. Both ends have a giant. The middle is open: the Fortune 5000, mid-size companies that can be worth $500m, many in bilingual markets where LLMs still write Hebrew any native can spot.
51 tokens in English, 127 in Hebrew. Same brief, same model. 12 languages on Claude: Spanish 1.4x, French 1.6x, German 2.0x, Arabic 2.4x, Hindi 3.4x. If your agents serve customers outside English, budget tokens per language. https://t.co/xZaQ7OHt67
Same model, same task, up to 5x the cost. Berkeley ran 7 models through 3 coding harnesses: success moved about 2 points, cost up to 5x. The harness is a pricing decision. In legacy industries it is also the permission layer, and no model swallows that. https://t.co/k8g4t9Phek