AI just reached half the world's population - and public trust in it just hit an all-time low.
•Mike Robinson
Use Left/Right to seek, Home/End to jump to start or end. Hold shift to jump forward or backward.
0:00
|
8:14
Yesterday in AI | Wednesday, April 15, 2026
Stanford just published its annual AI Index — and the trust numbers should make every AI executive uncomfortable. Meanwhile, an internal OpenAI memo leaked that reads more like a competitor takedown than an internal strategy doc. Anthropic is quietly building something that could blindside a $6.6 billion startup, NVIDIA dropped two open-source models on the same day, and Amazon just wrote a check that signals it's serious about taking on Starlink. It's a dense one.
Feedback? Email mike@yesterdayinai.news or connect on LinkedIn, X, Bluesky, or Substack. If you like the show, please take a minute to rate and review it so others can find it!
Hi folks, this is Yesterday in AI, your daily digest of everything happening in the world of artificial intelligence. I'm Mike Robinson. It's Wednesday, April 15th, and today's AI news reads like a company's internal Slack channel got accidentally made public, because in one case, it basically did. Let's get into it. Let's start with the Stanford 2026 AI Index, which dropped Tuesday and is one of the most revealing documents that research community publishes all year. The headline numbers are staggering. AI reached more than half of the world's population faster than the personal computer and the internet. Adoption metrics across virtually every sector are at all-time highs, and yet public trust is at all-time lows. Only 31% of Americans say they trust the government to manage AI responsibly. And here's the data point that stopped me in my tracks. 73% of AI experts say they feel optimistic about AI's impact on employment. Only 23% of the general public agrees with them. That 50-point gap is the widest expert public divide Stanford has ever recorded on this question. To put it plainly, the people building this technology think it's going to be fine. The people who are actually going to have their jobs disrupted by it do not. And there's a data point to sharpen that concern. Employment for junior developers between the ages of 22 and 25 has fallen 20% since 2024. That's not a projection. That's already happening. The AI index has become one of the most important annual documents for understanding where this technology is actually going versus where its advocates say it's going, and this year's addition lands at a moment when the credibility gap between those two things is as large as it's ever been. Now to the internal memo heard round the industry. The Verge published what appears to be an internal memo from OpenAI's chief revenue officer, Denise Dresser, and if you wanted evidence that OpenAI has shifted from ignoring Anthropic to being genuinely nervous about them, this document is it. Dresser calls Anthropic a single product company, argues that Anthropic's $30 billion annual revenue run rate is overstated by roughly $8 billion, and flags what she describes as a serious compute shortage on Anthropic's end as a strategic vulnerability. She highlights the staggering demand OpenAI has seen since its Amazon bedrock deal launched in February and frames this as evidence that OpenAI is pulling away. A few things are worth noting here. First, this is a CRO writing an internal memo about a competitor in terms that read like an investor pitch deck. Second, the fact that it leaked at all and that The Verge was able to confirm it as authentic suggest OpenAI's internal communications culture has some gaps. Third, when your head of revenue is putting competitor attack lines in writing inside your own company, that's usually a sign that the competitive threat feels existential rather than theoretical. The ironically titled Internal Memo That Became External News is not a great look for a company already navigating a rough stretch of press coverage. Let me turn to the agentic workplace race, because Tuesday brought three separate moves in the same direction. Microsoft is testing what they're calling OpenClaw style agents inside Microsoft 365 Copilot, always on AI that monitors your Outlook inbox, manages your calendar, generates a daily task list without being asked, and includes role-specific agents for sales, marketing, and accounting, each operating within defined permission boundaries. The testing is headed by Omar Shaheen, a former lead on the Word team. A public demo is planned for Microsoft's build conference on June 2nd. You may want to temper your expectations though. Google isn't sitting still either. The company quietly expanded its desktop agent inside Gemini Enterprise on Tuesday, adding a require human review toggle, a deliberate nod toward the agency with oversight model that Claude Cowork has been developing. The toggle is a small thing, but it signals that Google sees human-in-the-loop controls as a competitive feature rather than a limitation. The message: autonomous AI at your desk is the direction everyone is moving. The differentiator is going to be how much control users have over when it acts versus when it asks. And then there's the bombshell buried in the screenshot. Anthropic is reportedly building a vibecoding tool. Leak screenshots show Claude with a new in-chat app builder that can generate chatbots, photo albums, and landing pages directly from conversation. This is a direct shot at Lovable, the no-code app builder that was valued at 6.6 billion earlier this year. Anthropic building this capability into Claude's main interface rather than as a separate product is a meaningful strategic choice. It means they're betting that the people who already trust Claude for writing and reasoning are the same market that wants to build simple apps without a developer. Whether that's right is an open question, but the competitive implication is clear. Lovable just found out their primary differentiator may be getting absorbed into the world's most popular AI assistant. On the acquisition front, OpenAI has bought Hero Finance, a personal finance app that launched only five months ago. The app is shutting down April 20th, and the founding team is joining OpenAI. The founder, Ethan Block, is a serial entrepreneur whose previous exits include Digit, which sold for roughly $230 million, and Flowtown, which sold for $4.5 million. He has publicly noted that Hero was his 15th project, and that the first 13 didn't work. The strategic logic for OpenAI is straightforward. ChatGPT has been pushing into business finance use cases, and bringing in a team that has shipped two finance products gives them domain expertise they don't currently have. The AquaHire also tells you something about how OpenAI is thinking about the consumer finance space. They want people who've actually built in it, not just engineers who can read the API docs. Nvidia had a busy Tuesday on the research side. The company announced ISING, a family of open source quantum AI models designed for quantum computing calibration and error correction. The ISING calibration model can reduce quantum processor calibration time from days to hours. The ISING decoding model is two and a half times faster and three times more accurate than the PIMatching baseline, the previous standard for error correction. This is a relatively niche corner of the AI world, but it matters because quantum computing's practical progress has been largely bottlenecked by the manual, time-intensive calibration process. If AI can automate that, the development timeline for practical quantum computing compresses meaningfully. Nvidia making these models open source is consistent with their strategy of building goodwill in the research community while selling the hardware those researchers use. NVIDIA also released Audio Flamingo Next, an open large audio language model trained on a million hours of audio. The model comes in three variants, one for QA, one for reasoning with a novel temporal chain of thought approach that ties reasoning steps to specific timestamps, and one for audio captioning. On Long Audio Bench, Audio Flamingo Next scored 73.9 compared to Gemini 2.5 Pro's 60.4. That's a significant margin over Google's best audio model on a rigorous benchmark, and it's being released as open weights. For anyone building voice first or audio processing applications, this is a notable new baseline to build against. One more story before we close, and it's a big infrastructure move that didn't get the attention it deserved. Amazon has agreed to acquire GlobalStar's entire low Earth orbit satellite network for $11.57 billion. The deal gives Amazon a 200 satellite constellation, which is a meaningful footprint, though notably smaller than Starlink's 10,000 satellites. A key detail, Amazon explicitly committed to maintaining GlobalStar's existing iPhone and Apple Watch satellite emergency service contracts, meaning Apple users won't feel any disruption. The underlying strategy is clear. Amazon's Project Kuiper initiative has been moving relatively slow, and Starlink is pulling away from the field in direct-to-device connectivity. This acquisition is a shortcut. Whether 200 satellites is enough to mount a serious challenge to Starlink's 10,000 is a fair question, but 11.57 billion suggest Amazon isn't planning to lose this race without putting up a real fight. Stay curious, and I'll see you tomorrow.
