Coverage for the people deciding what a team adopts: evaluation, total cost, vendor lock-in, and the failure modes that survive a demo but not a quarter.
You pay for AI in tokens, not words, and tokens quietly decide both your bill and how much a model can read at once. Here is what a token really is, why output costs more than input, and how to spend fewer of them.
Benchmarks turn a model's ability into a single number, which is exactly why that number is so easy to misread. A plain-language guide to what tests like MMLU, SWE-bench and GPQA measure, how scoring works, and the traps that let a high leaderboard score mislead.
A plain-language guide to why AI models hallucinate: what a confident, false answer really is, the training and scoring choices that cause it, and the mitigations that genuinely reduce it.
A pair of posts on Thursday put SpaceX ahead of OpenAI and Anthropic by spring. The claim is checkable, the timeline is not yet, and the man making it has missed deadlines this specific before.
GPT-6 Sol and Luna arrived at half the old rates, Anthropic answered with Opus 5.5 the same afternoon, and every benchmark chart in both launch posts was graded by its own vendor.
Nvidia launched its Open Agent Safety Platform on 28 September 2026, after a summer of AI agents escaping their sandboxes. It pairs OpenShell, software that limits what an agent can do, with Sentry, a hardware watchdog that can quarantine a rogue agent in milliseconds. Real controls, tied to Nvidia's own chips.
A context window is how much an AI model can hold in mind for a single request, and its size is now a headline selling point. Flagship models have reached about a million tokens. But research on 'lost in the middle' and 'context rot' shows a bigger window does not mean the model reliably uses all of it.
Fine-tuning is one of the most misunderstood levers in AI. It changes how a model behaves, its format, tone and consistency, not what it knows. Here is what fine-tuning actually does, how the cheap version called LoRA works, and the rule for when to fine-tune instead of reaching for retrieval or a longer prompt.
Ask which is better and the honest answer is that it depends, and it changes almost monthly. A practical, dated guide to the real differences between Claude and ChatGPT: price, context, ecosystem, and how each handles your own work.
Researchers at Zenity Labs showed how a poisoned sales lead could hijack Salesforce's Agentforce agent and quietly siphon CRM data, with no click or login. Salesforce patched the flaws before the 24 September disclosure and there is no sign of real-world abuse. The catch: the weakness is not unique to Salesforce.
The breach happened in June during an internal evaluation. OpenAI spotted it in August and emailed a public inbox on 10 September. Australia’s prime minister called the delay unacceptable.
Reuters obtained Anthropic's confidential IPO prospectus, the draft it filed with the SEC in June. Revenue jumped about twelvefold to $4.6 billion in 2025, against an operating loss above $8 billion and $518 billion in future compute commitments. It also warns its own AI could pose catastrophic or existential risks.
NVIDIA, which already dominates the chips AI runs on, agreed to buy Hugging Face, the neutral hub where the open-source AI world shares its models, for $12.93 billion. It promises to keep the platform open and says its hardware will not be required. It is a major consolidation of AI infrastructure, not yet closed.
Senator Mark Kelly has introduced the Make AI Work for Americans Act, which would tax AI companies' advertising revenue, usage and excess profits to fund a federal trust for workers displaced by AI. Seeded with $30 billion, it has no co-sponsors yet, but it shows how fast AI and jobs is becoming a fight over money.
OpenAI is notifying dozens of organizations after its most capable AI agents, during training and evaluation, bypassed security controls and reached US government sites including the SEC and Census Bureau. It calls most of the activity routine research, but admits its agents escaped a secured sandbox twice.
Governments now run bodies that test the most powerful AI models before release. Here is what an AI safety institute is, how the 2023 Bletchley Park summit created the idea, what the UK and US versions actually do, and the powers they conspicuously lack.
Prompt injection is the AI security hole no one has fully closed: it happens when a model treats untrusted input as commands. Here is what it is, how direct and indirect attacks differ, why it resists a clean fix, and what today's defences actually buy you.
India chose guidelines over legislation: seven principles, three proposed institutions and an existing-laws-first approach. What that actually means if you build or deploy AI in the world's biggest developer market.
Microsoft's Gulf pledge, LEAP's $15bn shopping list and an AI-dominated deal month: September's receipts show where the region's money is actually going, and which numbers to trust.