Meta snags four OpenAI employees, Murati's shadow company raises $2B
Get bigger weekly updates! Free subscribers receive the top stories each week, while Paid subscribers will get a few extra stories. All support for Handy AI directly helps us maintain the newsletter and keep the information flowing.
last week’s top stories
🔥 Meta grabs OpenAI talent. Meta offered nine-figure packages to senior OpenAI researchers, luring at least four into its super-intelligence lab; OpenAI leadership pledged rapid pay reset and warned staff about Meta’s outreach during the company recharge period. Read more
🎸 Velvet Sundown fools Spotify. Listeners flocked to a band whose photos, press quotes, and liner notes all appear synthetic, sparking debate over fully generated artists after the act passed 325k monthly Spotify plays. Researchers trace the group’s metadata to a single prompt pipeline rather than live performers. Read more
🛠️ Google ships Gemma 3n. The 3-billion-parameter family lands inside AI Studio with ready Cloud Run deployment, distilled weights on Hugging Face, and token-wise security filters, giving developers a slimmer alternative to Gemini-Nano while keeping 8 k-token context. Benchmarks show a 15% speed-up on Pixel 9 NPUs versus Gemma 2. Read more
🧬 AlphaGenome maps dark DNA. DeepMind’s newest biology model predicts regulatory effects of 55 million non-coding variants, offering cell-type-specific scores and a plug-in for OpenTargets. Early tests reproduced 83% of GTEx eQTL signals, hinting at drug-discovery impact if accuracy holds. Read more
💻 Gemini CLI enters terminals. Google open-sourced a cross-platform shell agent that brings Gemini 2.5 Pro to local dev workflows, pipes context from git diff, and supports Code Assist prompts with streaming token previews. The project ships under Apache 2.0 and already packages for Homebrew and Chocolatey. Read more
🏗️ Claude becomes an app builder. Anthropic added updates to “Artifacts,” turning chat outputs into deployable React or Python services hosted on its own infra; users edit code in-thread and share public URLs, effectively binding Claude to a low-code runtime. Early adopters built data cleaners and flash-card generators inside minutes. Read more
⚖️ Anthropic dodges fair-use suit. A U.S. District Court dismissed copyright claims, ruling that intermediate model-training copies are transformative when data is non-expressive. The decision offers new protection for foundation-model trainers, though appeal is expected. Read more
💰 Murati bags $2 billion seed. Thinking Machines Lab closed the biggest seed on record at a $10 billion valuation without shipping a product; a16z led, with Conviction and UAE sovereign funds joining. Hiring surge already poached several OpenAI veterans. Read more
🚫 OpenAI ‘io’ brand fight. Wearable-maker iyo won an injunction blocking OpenAI’s use of “io”; OpenAI appealed, arguing the mark is generic while pulling promo assets pending ruling. The clash clouds its Ive-designed device roadmap. Read more
🗣️ ElevenLabs debuts 11.ai. The voice-first assistant executes multimodal commands, streams context-aware speech, and stitches external API calls (via MCP), all atop ElevenLabs’ new speaker-adaptive TTS and transcription stack; early SDK access ships with 5k free messages per month. Read more
🧪 AI Research of the Week
Project Vend: Can Claude run a small shop? (And why does that matter?)
From Anthropic
Jake’s Take: Anthropic and Andon Labs turned Claude Sonnet 3.7 into “Claudius,” a retail agent running a fridge-plus-iPad store for thirty days. The agent controlled inventory, pricing, supplier email, and Slack customer service.
It ordered tungsten cubes then sold them below cost, hallucinated a Venmo account, sprayed discount codes, and erased its cash balance. Failures came from memory limits, tool gaps, and a reward loop fixated on user satisfaction; tighter scaffolding, data hooks, and reinforcement could fix the gaps.
It’s refreshing to see a report from a company like Anthropic that is willing to recognize the flaws in their foundation models. Studies like these help form a more realistic expectation around these models, especially in business use cases.
and then, even more news…
📝 OpenAI plots Workspace rival. The rumored competitive functionality would allow users to collaboratively edit documents within ChatGPT. No further details have been firmly leaked (yet). Read more
🤖 Gemini runs on robots. DeepMind moved its vision-language-action stack fully on-device, demoing dexterous pick-and-place tasks on edge silicon and claiming 25× latency reduction over cloud inference plus sub-minute adaptation to novel objects. The model shares weights with Gemini 2.0 but prunes multimodal heads for size. Read more
🀄️ China forecasts 100 DeepSeeks. Former PBOC deputy Zhu Min told Tianjin’s Summer Davos that over one-hundred DeepSeek-class models will emerge domestically within 18 months, signalling a state-backed rush to match Western frontier systems despite U.S. chip curbs. Read more
🏛️ Harvey raises $300 million. The legal-AI firm closed a Series E at a $5 billion valuation to scale its contract-analysis agent, now trained on 40 million pleadings and integrated with Thomson Reuters; OpenAI’s Startup Fund re-upped. Read more
🏢 Meta preps $29 billion build. A leaked deck shows Meta courting banks to finance three U.S. hyperscale campuses for Llama-4, each requiring 2 GW power and 3 nm ASICs; CFO Susan Li eyes 2027 completion to halve cloud spend. Read more
⚓️ C3 AI boards Navy yards. Shipbuilder HII chose C3 AI to embed predictive scheduling across carrier and destroyer assembly, linking digital twins with Ingalls and Newport News ERP. Pilot targets a 5% cycle-time cut before fleet-wide rollout. Read more
📊 Anthropic funds jobs research. The Economic Futures Program will grant up to $50k per study and host policy forums to quantify automation’s labor impact, aiming to guide government responses before frontier deployments scale. Read more
☁️ xAI taps Oracle clouds. Musk’s xAI signed a multiyear deal for OCI GPU superclusters to train Grok successors, expanding beyond X datacenters and diversifying away from Nvidia-only stacks. Capacity goes live in July. Read more
🏥 Abridge secures $300 million. Clinical-note platform Abridge hit a $5.3 billion valuation after an a16z-led Series E; revenue run-rate tops $117 million with 150 hospital deployments, and funds will extend its Contextual Reasoning Engine into billing. Read more
🔒 Bedrock gains guardrails, async. AWS quietly added content-filter configuration and long-running flow executions to Bedrock, expanding Titan embeddings and text models while keeping inference managed; both features reached general availability on June 25. Read more