Leaked Files Show Alpha School's AI Did More Harm Than Good

01Alpha School's AI Does 'More Harm Than Good,' Leaked Files Show

Parents pay up to $65,000 a year for Alpha School's AI-generated curriculum. Internal company documents say those lessons sometimes do "more harm than good," according to a 404 Media investigation published February 17. Former employees described students as "guinea pigs" in a system that replaced teachers with software.

The school had earned powerful backing before those documents surfaced. CNN profiled Alpha in late January, asking whether AI schooling was "the future of education." Education Secretary Linda McMahon visited Alpha's Austin campus. At a White House virtual event on AI, First Lady Melania Trump shared the stage with an Alpha student. Bill Ackman, the billionaire hedge fund manager and Trump supporter, called Alpha "the first truly breakthrough innovation in K-12 education" since KIPP Academy.

Co-founded by Stanford graduate MacKenzie Price and Trilogy Software founder Joe Liemandt, Alpha runs 13 campuses across five states. The model eliminates teachers entirely. Human "guides" earning six-figure salaries supervise students who spend two hours each morning on math, science, and reading through AI software. Each student gets 30 minutes of one-on-one time with a guide per week. The rest of the day goes to workshops. There is no homework and no textbooks.

404 Media's reporting extends beyond flawed pedagogy. Alpha scraped content from other online courses without permission to build its training data, the outlet found. Student webcam recordings, screen captures, and names sit in a Google Drive spreadsheet shared so broadly that anyone with the link can access it. That includes people who have left the company.

None of this has dented Alpha's growth. The chain started in Austin at $40,000 annual tuition. A San Francisco campus opened this academic year at nearly double that price. Palo Alto and East Bay locations are planned. The Trump administration's endorsements all predated publication of the leaked documents.

Alpha is not alone in scaling AI instruction. The day after 404 Media's report, OpenAI announced education partnerships in India aiming to deploy ChatGPT Edu to more than 100,000 students, faculty, and staff within a year. Alpha's own internal records suggest the technology is not ready for the classroom it is already in.

02Google Puts AI Music Generation Inside Gemini for 200 Million Users

The music industry spent the past month drawing lines around AI-generated content. Google just stepped over one.

DeepMind's Lyria 3 audio model is now a beta feature inside the Gemini app. Users can generate 30-second instrumental tracks from text prompts, images, or video clips. The rollout is global. No separate app, no waitlist for most users: music generation sits inside the same chatbot that 200 million people already use for homework and email drafts.

The tool produces instrumentals only, no vocals or lyrics. Users describe a mood, upload a photo, or feed in a video, and Lyria 3 returns a half-minute clip. Ars Technica characterized the output as "something like music." Google is positioning the feature for personal and creative use: soundtrack a birthday video, score a short film, prototype a vibe.

The launch lands weeks after rights holders escalated legal and political pressure over AI-generated creative content, with new actions targeting companies that train generative models on copyrighted material. Google chose this window to ship music generation not as a research demo or a standalone tool for professionals, but as a feature embedded in its flagship consumer AI app.

The 30-second cap functions as a built-in disclaimer. Google is calling this experimental. The limit also means output is too short to substitute for a licensed track on any commercial platform. Whether that restraint holds depends on user demand and how quickly Google extends the duration. Beta labels in consumer products have a history of becoming permanent features before the policy debate catches up.

Google has not disclosed what music Lyria 3 was trained on or what licensing agreements, if any, cover its training data. The company says generated tracks include a SynthID watermark for identification. That is an attribution mechanism, not a rights framework.

No major label or publishing group has commented publicly on the Gemini integration. The silence may not last.

03Meta Locks In Millions of Nvidia Chips as Benchmark Scores Converge

Four models, four companies, less than one percentage point apart. That's the top of the SWE-bench coding leaderboard in February 2026. Meta, in the same week, signed a multiyear agreement for millions of Nvidia processors spanning two GPU generations.

SWE-bench published a rare full leaderboard refresh this month using independently generated results, not scores self-reported by labs. On the Verified benchmark, Claude Opus 4.5 scored 80.9%, Claude Opus 4.6 hit 80.8%, MiniMax M2.5 reached 80.2%, and GPT-5.2 landed at 80.0%. Three companies sit inside a 0.9-point band. Chinese labs GLM-5 and Kimi K2.5 also placed in the top ten. As Simon Willison observed, the practical gap between first and fourth amounts to statistical noise.

Meta's Nvidia contract covers Blackwell and Rubin GPUs alongside Grace and Vera CPUs. It marks the first large-scale deployment of Nvidia's Grace processor as a standalone product, not bundled into a Superchip module. Nvidia only began selling Grace separately in January 2026. The contract spans multiple hardware generations, locking Meta into Nvidia silicon through at least the Rubin cycle.

Neither company disclosed a dollar figure. Analyst Ben Bajarin estimated the package at billions. The Register reported it could add tens of billions to Nvidia's revenue over the contract term. Zuckerberg has said Meta plans to nearly double its AI infrastructure spending in 2026, with total investment potentially reaching $135 billion.

The Grace and Vera CPUs signal where that money is aimed. Both processors are optimized for inference, not training. As Meta deploys Llama across its products, the bottleneck shifts from building frontier models to serving billions of daily requests at low latency.

Meta gives Llama away for free. When four companies land within a single benchmark point, a proprietary model advantage is hard to sustain. A multigeneration Nvidia contract buys something more durable: inference capacity that takes years and billions of dollars to replicate.