AI
53 posts
DeepSeek 首次拉北大署名開源推理框架 DSpark|高並發下單用戶提速 60%-85%,吞吐反而再漲 51%-661%
DeepSeek 首次拉北大署名開源推理框架 DSpark|高並發下單用戶提速 60%-85%,吞吐反而再漲 51%-661% DeepSeek 與北京大學在 6 月 27 日聯合放出推理加速框架 DSpark,相關論文
Apple A20: first 2nm Apple Silicon, DRAM moves off-stack for thermals, 96-bit LPDDR6
Leaks show Apple's A20 will be the first 2nm Apple Silicon, repackaging DRAM to the side of the SoC (WMCM) for thermals, and jumping to 96-bit LPDDR6.
iOS 27 Beta 2 firmware reveals «Baidu Visual Search» — Apple's first region-routed visual search backend
iOS 27 Beta 2 contains a SearchPartnerInferenceProvider component with localized strings naming «Baidu Visual Search» — Apple's first region-routed visual search backend, and Baidu is the only named partner so far.
Apple Silicon's first generation skip: M6 ships base-only, M6 Pro/Max cut to make room for M7
Bloomberg reports Apple will skip M6 Pro and M6 Max entirely, pushing pro-tier Mac chips to the M7 family in late 2027 — the first break in Apple Silicon's annual Pro/Max cadence since 2020.
Xbox's third price hike in two years: Series X hits $800 in August, 2TB SKU killed
Xbox prices rise globally August 1: +$100 on 512GB SKUs, +$150 on 1TB, and the 2TB model is discontinued. Series X has risen 60% from its launch price.
DeepSeek ends its «small team» era — doubles every department, closes first $7.4B external round
DeepSeek announces a company-wide doubling of staff, just as it closes its first external round: ~$7.4B raised at a $50B+ post-money valuation. Tencent reportedly takes ~20%.
curl 8.21.0 patches a record 18 CVEs — one hidden 25 years, most found by AI
curl 8.21.0 fixes 18 CVEs at once — a single-release record. One had been hiding since 2001, and at least 11 were found by AI models, including 6 from AISLE alone.
Samsung's $648B Korean AI bet — SK Hynix joins, 2040s expansion pulled into mid-2030s
Samsung will announce a ₩1,000T (~$648B) ten-year plan, with SK Hynix joining. Korean chip expansion planned for the 2040s is being pulled into the mid-2030s by AI memory demand.
Cook signals more Apple price hikes coming — memory chip costs up 4× in three quarters
24 hours after raising Mac/iPad prices, Tim Cook tells the Wall Street Journal more hikes are coming. Memory chip prices have risen 4× in three quarters; iPhone 18 (September) is next.
GPT-5.6 needs White House sign-off, customer by customer — first time a US administration vets a frontier model rollout
OpenAI's GPT-5.6 enters limited preview — but the first ~20 enterprise customers each need White House approval. OpenAI says publicly: «this shouldn't be the long-term default.»
Apple's first broad price hike in 40 years — Mac Studio Ultra +$1,300, iPhone untouched
Apple raises prices across Mac, iPad, Vision Pro on June 25, averaging +$246. Mac Studio Ultra alone gains $1,300. iPhone, Apple Watch, AirPods stay unchanged. Cook calls it «a once-in-a-century flood».
GitHub adds Pull Request rate limits — first time the platform throttles «contributions welcome»
GitHub now lets repos cap how many open PRs a non-write-access contributor can have. PR throughput rose 3.6× in four years — AI-generated PRs are the main driver.
Nobel laureate John Jumper leaves DeepMind for Anthropic — Google's second AI defection in a week
John Jumper — Nobel laureate, AlphaFold lead, DeepMind VP — joins Anthropic. Less than a week after Gemini co-lead Noam Shazeer left for OpenAI. Two crown jewels gone in seven days.
Octoverse 2025: TypeScript #1, 180M devs, LLM repos triple
GitHub's Octoverse 2025: 180M developers (+36M in a year), TypeScript overtakes Python and JS as #1 by contributors, LLM-using repos grew 178% year-on-year.
Anthropic overtakes OpenAI in enterprise share — May 41% vs 39.5%
Per Ramp's expense data, Anthropic's share of enterprise AI subscriptions hit 41% in May, edging past OpenAI's 39.5% — first reversal. Export controls became a tailwind.
AMD FSR 4.1 lands on RX 7000 in July — INT8 version claims quality parity with FP8
AMD's FSR 4.1 expands from RX 9000 to RX 7000 in July. The INT8 build claims quality parity with the FP8 version. RX 6000 and Steam Deck GPU come early 2027 with shader-based inference.
Microsoft asks AWS to help save GitHub — 1B → 14B code changes in a year
Microsoft is asking AWS for help keeping GitHub up. Code-change events go from 1B in 2025 to a projected 14B in 2026 — 14× — driven by AI coding agents. Eight outages already this year.
Oracle's free ARM tier cut in half — effective today, over-limit free instances power off
Oracle Cloud Always Free's ARM instance ceiling drops 50% today (June 15) — from 4 OCPU + 24GB to 2 OCPU + 12GB. Over-limit free instances are powered off. No transition period.
US orders Anthropic to disable Claude Fable 5 and Mythos 5 — flagship models pulled
Commerce Secretary Howard Lutnick orders Anthropic to block all foreign users from Fable 5 and Mythos 5 — including its own non-US employees. Anthropic pulls both flagship models globally.
MacBook Neo 2: 12GB isn't an upgrade — it's the A19 Pro's stock config
MacBook Neo 2 will reuse A19 Pro binned chips left over from iPhone 17 Pro — which means 12GB LPDDR5X is standard, since the A19 Pro doesn't ship with 8GB. This unlocks AFM 3 Core Advanced.
Starlink V3: the 100× isn't marketing — it's two 10×s multiplied
Musk's math on Starlink V3: 10× bandwidth per satellite, 10× more satellites, multiplies to 100×. Orbit drops from 550km to 350km, halving minimum latency.
Google cuts its top security research team — same week it pitches AI security tools
Two weeks into Google Cloud layoffs, the Threat Intelligence Group (GTIG) was cut on June 3 — the same week Google Cloud Next launched new AI security products. Mandiant is also affected.
Bot traffic crosses human traffic — a year earlier than Cloudflare predicted
Cloudflare CEO Matthew Prince: bot traffic crossed human traffic for the first time — 57.4% vs 42.6%. He had predicted 2027 only three months ago.
Apple x OpenAI honeymoon over — OpenAI lawyers up, ChatGPT loses Siri exclusivity
OpenAI's lawyers are studying legal options against Apple — possibly a breach-of-contract notice, not necessarily a lawsuit. iOS 27 will end ChatGPT's exclusivity in Siri; users can pick Gemini or Claude.
Google quietly buys app source code from devs — email never mentions AI
Google emails top Android developers with «confidential content invitations» — offering to buy source-code access, for «improving Google's developer tools». The email never mentions AI. The linked page does.
NVIDIA's N1X drops tomorrow — Qualcomm's Windows-on-Arm monopoly is ending
NVIDIA, Microsoft and Arm posted in sync: a Computex Taipei keynote from Jensen Huang on June 1 will unveil N1 and N1X chips. Dell XPS, Lenovo Legion 7, ASUS ProArt, MSI, and Surface are all queued.
IBM bets $5B on open-source security — and openly admits Mythos was the trigger
IBM and Red Hat announce Project Lightwell — a $5B, 20,000-engineer open-source security clearinghouse. CEO Arvind Krishna says explicitly: Anthropic's Mythos was the trigger.
Linux CIFSwitch — 18-year-old kernel LPE, PoC public, found by AI semantic graph analysis
Linux kernel CIFSwitch — an 18-year-old local privilege escalation, found via AI-assisted semantic graph analysis. PoC public. Major distros vulnerable, patch and isolate now.
Dead Economy Theory — HN today: 676 points, 859 comments
Two Wharton economists formalize the «AI Layoff Trap»: each firm rationally cuts staff, but collectively this destroys consumer demand. Block laid off nearly half its workforce — stock jumped 25%.
Claude Code Dynamic Workflows — hundreds of agents in parallel, quarter-sized work in days
Claude Code Dynamic Workflows ships: Claude itself writes the orchestration script, spawning hundreds of parallel subagents in a single session, self-verifying before returning. Bun used it to port 750K lines to Rust in 11 days.
macOS Tahoe 26.5 patches 69 holes — kernel-fix count a multi-year record
macOS Tahoe 26.5 patches 69 security issues — 9 in the kernel alone, a multi-year high. One kernel and one WebKit fix credit «Claude and Anthropic Research».
Google insists users love AI Search — DuckDuckGo's no-AI version sees 28% spike same week
Google's CEO claims «people love AI Mode». DuckDuckGo's no-AI search page noai.duckduckgo.com saw +22.7% weekly traffic the same week, with iOS App installs spiking 69.9% at peak.
Claude Opus 4.8 ships — same price, more honest, Fast Mode cost cut 3×
Claude Opus 4.8 ships with Anthropic's H-round announcement. Same API price, beats Opus 4.7 across coding, agents, reasoning. 4× more likely to flag code bugs honestly. Fast Mode unit cost cut 3×.
Anthropic raises $65B at $965B — overtakes OpenAI as world's most-valued AI startup
Anthropic closes a $65B H-round at a $965B post-money — overtaking OpenAI as the highest-valued AI startup. ARR hit $47B this month. Likely the last private round before IPO.
DeepSeek's valuation: $10B → $50B in three weeks, National Big Fund in talks
DeepSeek's valuation jumped four times in three weeks — from ~$10B in early April to potentially ~$50B by May 6. China's National Big Fund is in talks to lead — its first investment in a domestic LLM.
iOS 27 Siri's new brain: a Gemini-distilled on-device model, with Google Cloud fallback
iOS 27 Siri will run a small model distilled from Google's full Gemini for on-device queries; some requests will go to a licensed Gemini in Google Cloud. Apple pays Google ~$1B/year for it.
Amazon kills its AI leaderboard — staff inflated token use to climb
Amazon employees gamed internal AI leaderboards by burning tokens on pointless tasks — «Tokenmaxxing». Meta's «Claudeonomics» board logged 60 trillion tokens in 30 days. Uber blew its annual AI budget in four months.
NYT: one job still growing in the AI era — cybersecurity
NYT: in the AI era, cybersecurity is one of the few jobs still growing.
Anthropic Glasswing month 1 — 10,000+ critical bugs, humans can't patch fast enough
Anthropic Project Glasswing's first month: Claude Mythos Preview and ~50 partners found 10,000+ high/critical vulnerabilities in major systems software. Of ~6,202 disclosed, only 75 patched. Humans can't keep up.
Musk and Zuckerberg killed Trump's AI EO — even «voluntary review» was too much
Hours before signing, Trump shelved a voluntary 90-day federal AI safety review EO. Musk, Zuckerberg, and David Sacks called Trump directly — the accelerationist faction won even against voluntary review.
Tesla FSD finally launches in China — but rivals are years ahead
Tesla launches FSD in China after years of waiting — but as Level 2, hands-on-wheel. Xpeng Mpilot already holds 60.1% of China NOA share; Baidu Apollo Go runs commercial robotaxis.
AI cybersecurity arms race — Anthropic Mythos vs OpenAI Daybreak
Within six weeks, both Anthropic (Project Glasswing + Claude Mythos) and OpenAI (Daybreak) entered enterprise cybersecurity. Mozilla used Mythos to fix 271 bugs in Firefox.
Huang admits it: China AI chip market handed to Huawei
NVIDIA CEO Jensen Huang on CNBC: «Huawei is very, very strong.» Q2 outlook excludes any China revenue. China used to be ~20% of NVIDIA's data center business.
Anthropic in talks with Microsoft for Maia chips — pays SpaceX $1.5B monthly
Anthropic faces severe compute shortage. Q1 demand grew 80× annualized; plans accounted for 10×. Now in talks to rent Microsoft Maia servers, while paying SpaceX $1.25B/month for compute.
AMD pours $10B+ into Taiwan — Helios AI servers ship H2 2026
AMD commits over $10B to Taiwan's semiconductor ecosystem on May 21 — partnerships with ASE and SPIL for advanced packaging, and Helios rack-scale AI platform shipping H2 2026.
Saudi Arabia pivots — The Line slows, $40B bets on AI datacenters
Saudi Arabia's PIF confirms The Line (170km linear city) will slow significantly. New 2026–2030 strategy prioritizes AI infrastructure and domestic investment. HUMAIN commits $40B to AI.
Anthropic pushes into Europe — Milan office, international staff to triple
Anthropic opens a Milan office, following Paris and Munich. EMEA revenue is growing 9× annualized — fastest-growing region. International headcount target: triple again, including new data center hires across Europe.
Meta cuts 8,000 — record revenue same quarter, $145B for AI
Meta started laying off ~8,000 on May 20, plus 6,000 unfilled roles cancelled (14,000 total reduction). Same quarter: AI infra capex of $125–145B planned. Free cash flow to drop ~80% to $8.5B.
Tencent Music finalizes Ximalaya buyout — SAMR forces end to exclusive rights
Tencent Music's acquisition of Ximalaya closed on May 18 — total ~¥18.6B. China's SAMR attached five conditions, the core one being giving up all exclusive audio rights and existing exclusive deals.
SpaceX S-1 published — Musk himself listed as a risk factor
SpaceX's S-1 was made public this week. In the 330-page filing, Musk himself is listed as a risk factor. xAI is mentioned 356 times, Tesla 87, Boring Co. and Neuralink appear too.
NVIDIA Q1 — $81.6B revenue, +85% YoY, $80B buyback announced
NVIDIA's FY27 Q1: $81.6B revenue (+85% YoY), data center $75.2B (+92%), $58.3B net income. Board authorizes $80B buyback. Quarterly dividend up 25× to $0.25.
AMD Ryzen AI Max 400 — first x86 chip to run a 300B model locally
AMD's new Ryzen AI Max PRO 400 (Gorgon Halo) bumps unified memory from 128GB to 192GB, up to 160GB dynamically assignable as VRAM. First x86 chip to run 300B+ parameter LLMs locally.
OpenAI prepares for IPO — largest AI listing ever, target $850B
OpenAI plans a confidential SEC IPO filing this Friday, targeting September listing at $850B+ valuation. Goldman and Morgan Stanley underwriting. Musk's lawsuit dismissed as too late.