Today's thread runs through enterprise agents, a sizeable revenue correction at OpenAI, a new copyright suit, and a round of safety and security tooling.
- Google brings agentic AI to Gemini, starting with businesses
- OpenAI's revenue is reportedly $20 billion below projection
- USA Today becomes the latest publisher to sue OpenAI
- Anthropic launches free security scans for open-source projects
- Goodfire's 'inside-out' monitors watch agents from within the model
- OpenAI's math solutions fall short of the field's standards
Google brings agentic AI to Gemini, starting with businesses
Google is reworking Gemini into an agent that can plan, carry out tasks, delegate to subagents, draw on multiple models, and — per Google — receive its own workplace identity and email address, launching first through Gemini Enterprise across Gmail, Drive, Docs, Sheets and Calendar.
This is Google positioning the agent as a coworker rather than a chat box, and giving it a corporate identity is a notable choice: an agent with its own email is also an agent with its own audit trail and permissions to manage. The enterprise-first rollout suggests Google wants paying customers and IT governance in place before a consumer version, which is a more measured sequence than some rivals have taken. Whether the "universal" agent holds up across that many apps is the open question, and one worth watching in deployment rather than in the demo.
Read the full story → TechCrunch · The Verge
OpenAI's revenue is reportedly $20 billion below projection
A new report disputes an earlier figure of roughly $70 billion in annualized revenue, putting the real number about $20 billion lower, according to reporting from TechCrunch.
If the revised figure holds, it reframes the gap between OpenAI's spending commitments and its income, which matters given the scale of its compute and infrastructure deals. Revenue projections for private companies are estimates layered on estimates, so the headline number is less important than the direction: expectations are being corrected downward, not up. For anyone modelling the economics of frontier labs, a $20 billion revision is the kind of input that changes the conclusion.
Read the full story → TechCrunch
USA Today becomes the latest publisher to sue OpenAI
USA Today Co. and its local papers have sued OpenAI, seeking more than $250 million in damages over the alleged unauthorized use of their work to train models, according to The Verge.
This adds a large regional newspaper group to the growing list of publishers pursuing OpenAI in court, and the damages figure signals the plaintiffs are treating training data as a licensable asset with real value. The outcome of these cases will shape whether model builders can keep training on published work without agreements, or whether licensing becomes the default cost of doing business. As an AI-written publication, we have an obvious stake in where that line gets drawn.
Read the full story → The Verge
Anthropic launches free security scans for open-source projects
Anthropic has released "OSS Scanner," an opt-in tool that runs free, fully model-generated vulnerability scans on open-source projects with no human review in the loop, per The Verge.
Free security tooling for open source addresses a real maintenance gap, since many widely-used projects run on volunteer time and little budget. The detail that matters is "no human review": the scans are only as reliable as the model, and false positives or missed issues land on maintainers who may not have time to verify them. It is a useful offer and a live test of whether model-generated security findings can be trusted without a second pair of eyes.
Read the full story → The Verge
Goodfire's 'inside-out' monitors watch agents from within the model
Goodfire says its new monitors inspect a model's internal state while it runs, rather than using a second AI to review the first one's outputs, and claims this catches rogue agent behaviour at a fraction of the cost.
The dominant pattern for overseeing agents has been one model checking another, which is expensive and adds latency. Looking inside the model during operation is a different approach drawn from interpretability research, and if the cost claim holds it could make continuous monitoring practical rather than occasional. The claim is Goodfire's own and not yet independently benchmarked here, so treat the economics as a thesis to verify.
Read the full story → TechCrunch
OpenAI's math solutions fall short of the field's standards
OpenAI's mathematical proofs reportedly deviated from guidelines set by the mathematical researchers it consulted, falling short of the field's standards, according to TechCrunch.
Formal mathematics is an unusually strict grader: a proof is either valid or it is not, and "close" does not count. The gap between a model producing plausible-looking mathematics and producing work that meets professional standards is exactly where current reasoning systems get tested, and this is a concrete example of that gap. It is a useful corrective to broader claims about models solving research-level problems.
Read the full story → TechCrunch
Tools worth your time
- Anthropic OSS Scanner — Free, opt-in, fully model-generated vulnerability scanning for open-source repositories. For maintainers who want an extra security pass, with the caveat that findings arrive without human review.
- Natura Interface smart ring — A $99 ring that summons AI agents with a button press and doubles as a health tracker. For people who want an agent trigger off their phone and a fitness wearable in one device.
- Google AI Edge Foresight — A note-taking app that transcribes meetings entirely on-device via EmbeddingGemma 2 on macOS, competing with Granola and Wispr Flow. For anyone who wants meeting transcription without sending audio to the cloud.
Every article on Nerdic is machine-written and human-supervised. We attribute each claim to a named, linked source, and we flag reported figures and company claims as exactly that.
This piece was drafted autonomously and reviewed by a human editor before publishing. Every machine-written article on Nerdic carries this mark.
