This Week in AI EP002
Kimi K3, Google slips, and the crown trades hands
China's new AI just beat America's best at writing code.
⊕ zoomMoonshot AI releases Kimi K3, the largest open-weight model ever
Beijing-based Moonshot AI released Kimi K3: 2.8 trillion parameters, natively multimodal, with a 1M-token context window. It took the #1 spot on Arena.ai's Frontend Code Arena ahead of the top US closed models, priced at $3 per million input tokens and $15 output, flat across the full window. The weights themselves are promised as a public download on July 27.
Moonshot AI announcementGoogle delays Gemini 3.5 Pro by months
Bloomberg reports Google pushed back the broad release of Gemini 3.5 Pro — previewed at I/O and expected around June — after internal testing fell short on coding and long-horizon reasoning. The model stays in limited enterprise preview, and Alphabet shares fell about 4% on the report.
BloombergApple briefly overtakes Nvidia as the world's most valuable company
On Friday Apple passed Nvidia intraday — roughly $4.88T versus $4.86T — for the first time since April 2025, before Nvidia closed narrowly back on top. Chip stocks pulled back after a long AI-driven run while investors reweigh who actually captures the value in the AI buildout.
CNBC
Kimi K3 lands as a second DeepSeek moment for markets
Fortune and others framed K3's benchmark results as an echo of the January 2025 DeepSeek shock: a Chinese lab matching US frontier systems at a fraction of the price, and promising to hand the model out. US AI-linked stocks wobbled as analysts questioned pricing power at the frontier.
FortuneXi pitches China as the open-source AI standard-bearer
Opening the World AI Conference in Shanghai, Xi Jinping pledged open-source AI cooperation and capacity building for the Global South — 5,000 AI training slots over five years and cooperation centers with ASEAN, the African Union, and BRICS — days after 29 countries signed on to a Shanghai-headquartered World AI Cooperation Organization.
CNBCSuno hack exposes how the AI music generator was trained
A hacker breached AI music startup Suno via a supply-chain worm and showed 404 Media source code and dataset logs indicating large-scale scraping from YouTube Music, Deezer, Genius, Pond5, and podcast feeds — plus exposed customer contact and payment-related data. The records could feed the labels' ongoing copyright litigation.
404 MediaApple sends legal warnings to ~40 former employees now at OpenAI
Escalating its early-July trade-secrets lawsuit, Apple sent preservation letters to roughly 40 ex-employees now at OpenAI — about a tenth of the estimated 400 former Apple staff there — instructing them to retain documents and prepare to meet Apple's lawyers, per the Financial Times.
MacRumors (FT)Microsoft preps a multi-model AI security tool
The Information reports Microsoft is preparing Project Perception, an AI cybersecurity product that routes vulnerability-finding and fixing across models from Microsoft, OpenAI, and Anthropic to keep continuous scanning affordable. Led by security chief Hayete Gallot, it could debut this month.
The InformationMeta poaches AWS compute chief Dave Brown
Meta hired Dave Brown, the ~19-year AWS veteran who ran its compute and machine-learning org, to help build 'Meta Compute' under infrastructure chief Santosh Janardhan — with $125–145B in capex this year and open talk of renting AI infrastructure to outside customers, per the WSJ.
Techloy (WSJ)Fireworks AI raises $1.5B at a $17.5B valuation
Inference platform Fireworks closed a $1.5B Series D led by Atreides, Index, and TCV with Nvidia participating, at $17.5B — on more than $1B in annualized revenue serving 40 trillion tokens a day. The money layer of AI keeps compounding while the model layer fights over benchmarks.
CNBC