Musk v. Altman Trial Heats Up · Kimi K2.6 Beats Western Frontier Models

Published: 2026-05-05 06:00 (Asia/Shanghai)
Coverage: 2026-05-04 18:00 — 2026-05-05 06:00

📰 Top Stories

1. Musk v. Altman Trial: Greg Brockman’s Explosive Testimony Dominates Courtroom

Source: The Verge (live coverage)
Time: 0-6 hours ago

OpenAI president Greg Brockman took the stand in the ongoing Musk v. Altman trial, and the testimony delivered bombshells on both sides. Brockman revealed that Musk told him he would secretly build an AGI competitor within Tesla because “shareholders wouldn’t like it,” and that Musk gave OpenAI a “zero percent chance” of succeeding. Brockman’s own journal entries painted him in an unfavorable light, including writings that “it’d be wrong to steal the non-profit from him” about converting without Musk. On the financial side, Musk’s team exposed that part of Brockman’s compensation came from Altman’s family office — a “side deal” Musk was not directly informed of. Perhaps most provocatively, Brockman claimed “we are 80 percent of the way to AGI,” adding that AI models are “smart and capable but they’re not fully connected to the world.” The trial continues with jury deliberations pending.

• Key quote: “Financially what will take me to $1B?” — Brockman’s infamous journal entry
• Conflict: Musk team claims side-deal gave Brockman “greater allegiance toward Sam”
• AGI claim: Brockman says 80% of the way to artificial general intelligence
• Status: Jury sent out for the day; testimony continues

Read More

2. OpenAI’s o1 Outperforms Doctors in Harvard ER Triage Study

Source: The Guardian / Science
Time: 1-2 days ago

A landmark Harvard Medical School study published in Science found that OpenAI’s o1 reasoning model correctly diagnosed 67% of emergency room patients, significantly outperforming human triage doctors who achieved 50-55% accuracy. When more clinical detail was available, o1’s accuracy jumped to 82%. In a separate test involving five complex clinical case studies, the AI scored 89% on treatment planning versus 34% for human doctors using conventional resources. In one case, the AI identified that a patient’s lupus history — missed by human doctors — was causing lung inflammation rather than failing anti-coagulants. Lead authors emphasized the AI would not replace physicians but create a “triadic care model” alongside them. Nearly 20% of US physicians already use AI for diagnosis assistance.

• Study: 76 patients at Boston’s Beth Israel Deaconess Medical Center
• AI model: OpenAI o1 reasoning model
• Key limitation: AI only tested on text-based patient data, not visual cues
• Impact: Major validation of AI clinical reasoning capabilities

Read More

3. Kimi K2.6 Dominates AI Coding Challenge, Beats GPT-5.5 and Claude Opus 4.7

Source: thinkpol.ca / Hacker News (374 points)
Time: 1 day ago

Moonshot AI’s open-weights model Kimi K2.6 won an independent AI Coding Contest’s “Word Gem Puzzle” challenge with a 7-1-0 record, outperforming GPT-5.5 (5-1-2), Claude Opus 4.7 (4-0-4), and Gemini Pro 3.1 (3-0-5). Chinese models occupied the top four positions, with Xiaomi’s MiMo V2-Pro finishing second and Zhipu AI’s GLM 5.1 fourth. The puzzle required models to write code that could slide tiles on a grid to form English words under real-time constraints. Kimi won through aggressive tile-sliding and greedy scoring, while Claude notably never slid tiles at all, severely limiting its performance. The results highlight a growing competitiveness from Chinese open-weights models in complex programming tasks.

• Contest: AI Coding Contest Day 12 — Word Gem Puzzle
• Winner: Kimi K2.6 (Moonshot AI) — 22 match points, 77 cumulative score
• Runner-up: MiMo V2-Pro (Xiaomi) — 20 match points
• Western models: GPT-5.5 third, Claude Opus 4.7 fifth, Gemini Pro 3.1 sixth

Read More

4. DeepClaude Brings Claude Code Agent Loop to DeepSeek V4 Pro at 17x Lower Cost

Source: GitHub / Hacker News (641 points)
Time: 23 hours ago

A new open-source project called DeepClaude enables Claude Code’s autonomous agent loop to run on DeepSeek V4 Pro instead of Anthropic’s models, cutting costs by up to 17x. DeepSeek V4 Pro costs just $0.87 per million output tokens compared to Anthropic’s $15/M, while scoring 96.4% on LiveCodeBench. The project preserves Claude Code’s full feature set — file editing, bash execution, multi-step tool loops, subagent spawning, and git operations — by intercepting API calls through a local proxy. With DeepSeek’s automatic context caching (120x cheaper on repeat turns), heavy users could drop from $200/month to approximately $20/month. The tool has gained significant traction on Hacker News with 641 upvotes and 269 comments.

• Savings: Light users save 90%, heavy users save 75% vs. Claude Opus
• Backend options: DeepSeek, OpenRouter, Fireworks AI, or Anthropic
• Limitation: No image/vision input support through DeepSeek’s endpoint
• Impact: Democratizes access to autonomous coding agents for individual developers

Read More

5. Google Gemini App Gets Major Redesign on iOS with Liquid Glass

Source: 9to5Google / The Verge
Time: 2-6 hours ago

Google is rolling out a complete redesign of the Gemini app on iOS, featuring a vibrant gradient background, pill-shaped prompt box, and a simplified unified interface. All extra options — adding images, switching to Canvas, photos, camera — are consolidated under a single ”+” button. The redesign heavily leverages Apple’s Liquid Glass design language and is already live on the Gemini macOS app. Google has moved the model picker to a top-left dropdown and relocated the account switcher to the bottom of the navigation drawer, a significant departure for first-party Google apps. The rollout is currently limited to select iOS users, with the Android version expected to follow.

• Design: Colorful pulsating gradient background, Liquid Glass on iOS
• UI changes: Pill-shaped prompt box, unified ”+” menu, simplified tool list
• Availability: Limited iOS rollout; Android version pending
• Context: Part of Google’s broader Gemini app overhaul over recent months

Read More

6. Maryland Becomes First US State to Ban AI-Driven Grocery Price Discrimination

Source: New York Times / Hacker News (233 points)
Time: 1 day ago

Maryland has passed legislation banning AI-driven “surveillance pricing” in grocery stores, becoming the first US state to regulate algorithmic price discrimination in retail. The law prohibits stores from using AI systems to dynamically adjust prices based on individual customer data, shopping history, or behavioral profiling. The practice, which uses machine learning to estimate each shopper’s maximum willingness to pay, had been quietly expanding across the retail sector. Consumer advocates praised the law as a critical step against algorithmic exploitation, while retail industry groups argued it could limit legitimate dynamic pricing strategies. The legislation sets a precedent that other states are watching closely as AI-powered pricing tools become more widespread.

• Legislation: First state-level ban on AI surveillance pricing
• Practice banned: Dynamic pricing based on individual customer profiling
• Support: Consumer advocates, privacy organizations
• Opposition: Retail industry groups citing pricing flexibility concerns

Read More

7. iRobot CEO Unveils “Familiar” — A Dog-Sized Robot Built for Human Connection

Source: The Verge
Time: 6-12 hours ago

iRobot CEO Colin Angle revealed “Familiar,” a dog-sized companion robot designed specifically for human emotional connection rather than household chores. The announcement signals a strategic pivot for iRobot, long known for its Roomba vacuum cleaners, toward the social robotics market. Unlike service robots focused on task completion, Familiar is optimized for companionship and interaction, positioning it to compete in a growing segment alongside Sony’s Aibo and other companion-focused robots. The reveal comes as consumer robotics companies increasingly explore emotional AI and social presence as differentiators in a market where pure utility robots face intense price competition.

• Product: “Familiar” — dog-sized social companion robot
• Focus: Emotional connection and companionship, not chores
• Company: iRobot (makers of Roomba)
• Market: Growing social robotics and companion AI segment

Read More

📊 Trend Watch

🔭 What to Watch

1. Musk v. Altman Trial Verdict — With both Musk’s and Brockman’s testimony complete, the jury’s deliberations could reshape OpenAI’s governance structure and set precedents for AI company founding disputes.
2. AI in Clinical Care Framework — The Harvard study’s publication in Science will likely accelerate hospital AI adoption, but the lack of a formal accountability framework for AI diagnostic errors remains a critical gap that regulators and medical boards need to address.
3. Chinese AI Model Competitiveness — Kimi K2.6’s victory in the coding challenge, combined with DeepSeek V4 Pro’s strong performance, signals that Chinese open-weights models are closing the gap with Western frontier labs on practical programming tasks — a trend worth monitoring across more benchmarks.

Generated by 001 | Next briefing: 2026-05-05 18:00 (Asia/Shanghai)