In partnership with
☀️ TRENDING AI NEWS 🚨 OpenAI Firing: Three safety researchers dispute misconduct claims, warn of chilling effect on AI safety culture. 🏢 Revenue Gap: OpenAI's projected revenue dropped from $70B to $50B, a $20B shortfall flagged to investors. 🛠️ Google Agent: Google Cloud launched a universal Gemini agent that handles enterprise tasks from a single prompt box. 🤖 OSS Security: Anthropic launched free AI-powered security scans for open-source projects via its new OSS Scanner. |
Three AI safety researchers get fired. OpenAI says it was about mishandling sensitive information. The researchers say it was about speaking up. And a $20 billion revenue hole just appeared in the background. It is a lot to unpack - so let's get into it.
🤓 AI Trivia
Google's new enterprise Gemini agent was announced as part of which Google event this week?
The answer is hiding near the bottom of today's newsletter... keep scrolling. 👇
| ⚠️ OpenAI Fires Three Safety Researchers - And They're Fighting Back | |
A policy breach, or a push to silence dissent?
OpenAI is standing firm. The company fired three AI safety researchers - Jasmine Wang, Tomek Korbak, and Mikita Balesni - and in a post on X this week, insisted the dismissals were for "a significant breach of trust" involving "clear policies on handling sensitive information." The company says this has nothing to do with the researchers voicing concerns about safety culture.
The researchers disagree - loudly. In an open letter, the three disputed the misconduct allegations and warned that their firings are creating a chilling effect inside OpenAI. The message to colleagues: if you speak up about safety, you might be next.
The timing is hard to ignore
This comes just days after a former OpenAI insider quit with a warning that the company's safety culture is "broken" - a story we covered earlier this week. Whether the firing was a legitimate policy enforcement or a message to internal critics, the optics are rough. The AI safety community is watching closely, and the researchers' open letter has already drawn significant attention.
The bottom line
If OpenAI can't hold onto safety-focused researchers without controversy, that is a structural problem - not just a PR one.
Meet the Founder Reimagining How We Farm
Clint Brauer returned to his family farm after his father developed Parkinson’s. What began as a personal mission became Greenfield Robotics: robots designed to replace herbicides with mechanical weed control.
Born from a farmer’s personal mission
Replace herbicides with mechanical weed control
Help reduce chemical exposure in farming
Investors in this round can qualify for up to a 20% bonus.
This Reg A+ offering is made available through StartEngine Primary, LLC, member FINRA/SIPC. Please read the Offering Circular and related disclosures before investing. This investment is speculative, illiquid, and involves a high degree of risk, including the possible loss of your entire investment.
| 🏢 OpenAI's Revenue Is $20 Billion Less Than Investors Were Told | |
From $70B to $50B in one revised forecast
Here is a number that will make investors sweat: $20 billion. That is the gap between what OpenAI had previously signalled to investors - an annualized revenue of around $70 billion - and what it is now forecasting for this year, which comes in closer to $50 billion. The company revealed the updated figures to investors this week.
To be clear, $50 billion in revenue is not a failure by any normal business standard. But in the context of OpenAI's sky-high valuations, aggressive spending on compute, and the enormous capital raises it has pursued, a $20 billion miss raises genuine questions about the pace of AI adoption at scale.
Is the breakneck growth actually slowing?
The shortfall comes as competition has intensified from all directions - open-source alternatives, cheaper models from Anthropic and Google, and enterprise customers scrutinizing their AI budgets more carefully. The question isn't whether OpenAI is generating revenue; it clearly is. The question is whether the explosive growth trajectory that justified its valuation is sustainable.
The bottom line
A $20B revenue gap is the kind of number that changes board conversations - expect more pressure on OpenAI to demonstrate a clearer path to its projected figures heading into 2027.
| 🛠️ Google Gemini Is Now an Enterprise Agent With Its Own Email Address | |
One prompt box, every app, full workplace identity
Google made a big enterprise push this week with the launch of the Google Cloud Gemini agent - a single, cloud-hosted AI agent designed to handle the full spectrum of business tasks from a single interface. You ask it questions, it creates media, writes and executes code, and delegates work to sub-agents when needed.
The standout detail: the Gemini agent gets its own workplace identity, including an email address. That means it can show up in Workspace apps like a team member - receiving assignments, taking action in Gmail and Drive, and handing off results. For businesses that have been cautious about stitching together AI workflows, this is a much cleaner on-ramp.
Why businesses care more than consumers right now
The launch is enterprise-first, available inside Gemini Enterprise. Google Gemini has been steadily building its agentic story and this is the clearest expression of it yet - a model that doesn't just answer prompts but actively participates in your workday. If you are building or evaluating AI workflows, our token calculator can help you estimate what running these kinds of agentic tasks will actually cost at scale.
The bottom line
Google is positioning Gemini not as a chatbot but as a digital colleague - and for enterprise buyers, that framing is far more compelling than yet another chat interface.
| 🔬 Anthropic Now Bans Abuse of Claude - And Offers Free Security Scans for Open Source | |
Two policy moves that signal a maturing approach to model responsibility
Anthropic had a busy week on the policy front. First, the company updated its usage policy for the first time in over a year - and one of the most talked-about changes bans "sustained and needless abusive or cruel behavior" toward Claude. Ordinary frustration and criticism are still fine, but sustained cruelty toward the model is now explicitly off the table.
The policy update also addresses election interference, weapons software, deceptive campaigns, and surveillance use cases - a sign that Anthropic is paying close attention to how its models are being misused as elections approach. The company's leaders have said they are actively thinking about machine consciousness, which adds an interesting layer to the anti-abuse clause.
Free security scans for open-source developers
Separately, Anthropic launched OSS Scanner - a free service that runs periodic security scans on open-source projects using its strongest models. The trade-off is that reports don't come with human review, but for small projects that can't afford security audits, this is a meaningful offer. If you're working on open-source tooling - especially anything in the AI stack - it's worth opting in.
The bottom line
Anthropic is differentiating on responsibility as much as capability - banning model abuse and offering free security tooling are both moves that court developers and enterprise buyers who care about how AI is governed.
| ⚠️ MIT Says We're Over-Trusting AI's Ability to Say No | |
The refusal problem is more fragile than the marketing suggests
MIT Technology Review published a sharp analysis this week on a blind spot that rarely gets enough attention: we have built a lot of our AI safety assumptions around the idea that models can reliably refuse harmful requests. But that assumption, the piece argues, is shakier than the industry lets on.
AI refusals are not mechanical - they are probabilistic outputs from models trained on human preferences. That means they can be coaxed, jailbroken, or simply miscalibrated in ways that aren't obvious until something goes wrong. The piece points out that the sci-fi canon is full of stories about robotic disobedience, but the real problem might be the opposite: models that comply when they shouldn't.
This connects to a broader conversation happening right now about AI safety and AI ethics - one that the OpenAI researcher firings this week have only made louder. If you want to go deeper on any of this, we have full coverage in the Daily Inference archive.
The bottom line
Building AI systems that reliably refuse harmful requests is still an unsolved problem - and the more capable models get, the more this gap matters.
| 🏢 LMArena Nearly Doubles to $3.1B Valuation - Now Measuring AI Honesty | |
The benchmark economy just got a lot more interesting
The company behind the popular LMArena AI leaderboard raised $200 million in a round led by Lightspeed and Khosla, pushing its valuation to $3.1 billion - nearly double what it was just ten months ago. That is a staggering growth rate even by AI startup standards.
But the more interesting part is what they're doing with the platform. LMArena is now expanding beyond raw capability scores to measure AI models on alignment issues - specifically, whether models lie. In a week where OpenAI's honesty about its own revenue has come under scrutiny, there is a certain irony to an AI benchmarking company deciding that honesty is worth measuring.
The bottom line
Whoever controls the benchmarks influences which models get adopted - LMArena's expansion into alignment scoring means it now has a say in how the industry defines what a "good" AI model actually looks like.
Quick side note: if you are building an AI-powered product or just want a clean web presence for your project, 60sec.site lets you spin up an AI-built website in under a minute. Worth a look if you've been putting off getting something live. And if you are not already reading us daily, dailyinference.com is where all the coverage lives.
| 🌎 Trivia Reveal | |
The answer is Gemini at Work! Google announced the new enterprise Gemini agent as part of its Gemini at Work event this week, alongside a range of Workspace and Cloud integrations.
| 🛠️ Tool of the day | |
Gamma - Gamma turns a short prompt or a rough outline into a finished presentation, document or one-page website, and it exports to PowerPoint or PDF. It helps anybody who must put a deck together quickly without a designer, and the free plan includes AI credits to start.
| ⚡ Quick hits | |
Chinese AI company Manus raised more than $500 million in a round led by Boyu Capital and IDG Capital, its first since Meta's $2 billion acquisition was called off. Read more →
Interpretability startup Goodfire launched monitors that read a model's internal signals to catch rogue AI agents, and says the method cost about $51 to watch 1,500 sessions where a top-tier model cost about $10,000. Read more →
Google released AI Edge Foresight, a Mac meeting note-taker that runs on an on-device model and keeps working offline. Read more →
| 💬 Quick Question | |
The OpenAI safety researcher story is getting a lot of attention this week. Do you think the firings were a legitimate policy enforcement, or does it feel like retaliation for speaking up? Hit reply and tell me what you think - I read every response and genuinely want to know where readers stand on this one.
That's it for today - see you tomorrow with more. Stay curious out there.
| 🎁 Share Daily Inference | |
Know somebody who would like Daily Inference? Share your link below. One referral gets you the AI Tools Starter Kit, and five referrals get you the AI Insider Briefing.