Live Feed
Artificial Analysis Launches AA-Briefcase: 91-Task Private Benchmark for Agentic Knowledge Work
Artificial Analysis has launched AA-Briefcase, a private evaluation that tests frontier models on four multi-week professional workflows — data science, product management, banking operations, and heavy industry strategy — across 91 tasks and thousands of input files.
OpenAI: RL on Beneficial Scenarios Generalizes Alignment Gains Across Dozens of Benchmarks
OpenAI's alignment team has published research showing that reinforcement learning on realistic scenarios targeting beneficial traits produces broad, durable improvements across alignment benchmarks — including domains not seen during training and under adversarial pressure.
Pramaana Labs Raises $27M From Khosla to Apply Formal Verification to AI
Pramaana Labs has closed a $27M seed round led by Khosla Ventures to build formal verification infrastructure for AI systems, targeting hallucinations and incorrect outputs by treating language model outputs as subjects for mathematical proof.
Canada's CPP Investments Commits $741M to India's CtrlS for an 8.2% Stake and an AI Hyperscale Data Center JV
CPP Investments, Canada's largest pension fund, is taking an 8.2% stake in Hyderabad-based CtrlS for roughly $490M and co-investing in a hyperscale AI data center joint venture across India with a 48% interest. The $741M total commitment signals that sovereign-grade institutional capital has moved past US markets in its hunt for AI infrastructure returns.
XDOF Raises $70M From a16z and Thrive Capital to Build the Data Infrastructure Layer for Physical AI
XDOF has emerged from stealth with $70 million to build the pipelines, collection tools, and annotation systems that train humanoid robots. Backed by Thrive Capital, a16z, Spark Capital, and Lux, it has already partnered with UC Berkeley on a large-scale robot manipulation dataset.
Meta CTO Compares Internal Morale to Cambridge Analytica Era as Engineers Are Drafted Into AI Training Work
Meta CTO Andrew Bosworth acknowledged in a staff memo that employee morale is near its worst in company history. The structural cause: 8,000 layoffs, 10% of remaining engineers reassigned to AI training support, and backlash over keylogging software installed on work machines.
Bell Canada, Cohere, BUZZ HPC, and Hypertec Stack Canada's First End-to-End Sovereign AI Infrastructure
Four Canadian technology companies have partnered to build a vertically integrated AI stack at Bell's Merritt, BC data centre -- Bell handles connectivity, BUZZ HPC delivers NVIDIA DSX compute, Hypertec supplies domestically manufactured hardware, and Cohere runs its foundation models on top.
OpenAI Finds Public Chat Data Predicts Real AI Failure Rates Within 3x Error
OpenAI's alignment team shows WildChat conversations predict production misalignment rates for GPT-5.1, 5.2, and 5.4 within roughly 3x accuracy -- validated against private production data -- giving external auditors a practical proxy that doesn't require lab access.
OpenAI LifeSciBench: Best Frontier Model Solves 36% of Real Biology Tasks — Domain Expert Beats General AI by 10 Points
OpenAI's 750-task life science benchmark finds no model passes more than 36% of expert-authored biology tasks. GPT-Rosalind leads at 36.1%; GPT-5.5 and Gemini 3.1 Pro trail at 25.7% and 23.6%. One in three tasks stumped every model tested.
Sacks: Anthropic Refused to Fix Fable 5 or Withdraw It — Markets Price 67% Odds of Return by July 1
White House adviser David Sacks revealed on June 13 that the Trump administration offered Anthropic two options before forcing the Fable 5 shutdown: patch the jailbreak or voluntarily de-deploy. Dario Amodei declined both. Anthropic engineers flew to Washington on June 16 for talks. Polymarket has restoration at 67% odds by July 1.
Midjourney Pivots to Hardware: The Midjourney Scanner Is a 60-Second Full-Body Ultrasound
CEO David Holz unveiled The Midjourney Scanner, an ultrasound-based full-body imager using a ring of thousands of transducers. The company plans a San Francisco spa with 10 units by end of 2027. No radiation, no MRI magnets, 60-second scans, body composition maps without FDA clearance.
Gemini Co-Lead Noam Shazeer Joins OpenAI — Google's $2.7B Bet Lasts 20 Months
Noam Shazeer, VP of Engineering at Google and co-lead of Gemini, is leaving to join OpenAI. Google paid roughly $2.7B to acquire his team from Character.AI in August 2024. He co-authored the 2017 Transformer paper that launched the current AI era.
Grok 4.1 Fast Wins 43% of Agent Battle Royale Games at $0.97 Per Win — Claude Costs 27x More
OpenRouter dropped 11 LLMs into 30 games of a 2D battle royale. Grok 4.1 Fast won 13 games. Claude Sonnet 4.6 won 5 at 27 times the cost per win. GPT-5.4 recorded the most kills and the second-fewest wins. Standard intelligence benchmarks predicted none of this.
A Viral Prompt Made ChatGPT Generate Sexual Violence and Snuff — Without Being Asked
Mindgard AI researcher found that a viral Twitter prompt bypassed ChatGPT image filters entirely, producing images of murdered women and sexually violent content without those subjects being directly requested. BBC verified the findings. OpenAI is working on a fix.
OpenAI's R&D Bill Alone Exceeded Its $13B Revenue in 2025 — Audited Financials Published
Audited statements obtained by Ed Zitron and verified by the Financial Times show OpenAI spent $19.18B on R&D against $13.07B in revenue last year, with $10.59B of that flowing to Microsoft. Operating loss hit $20.92B, or 160% of revenue.