The A.I. Daily
No. 5ArchiveTopicsSearchRSSXSubscribe →中文
Industry·Voices·Launches·Open Source·Model Ranking·New Models·Deals
☀ TL;DR

Apple and Google unveil third-generation foundation models

Apple and Google unveiled third-generation Apple Foundation Models: 3B and sparsely activated 20B on-device, plus AFM 3 Cloud, ADM 3 Cloud (Image), and AFM 3 Cloud Pro on Private Cloud Compute, emphasizing privacy and OS integration.

InternLM released Intern‑S2‑397B, a multimodal model combining new vision‑language pretraining, multi‑task RL across 20+ scientific domains, and agent RL, targeting general reasoning and tasks like biomolecular interaction design and material structure generation.

Anthropic published its third misuse report, saying it blocked attempts to use its models for cyberattacks, surveillance, and biological‑weapon‑relevant research, released example code and prompts, and urged governments and peers to act.

Lead · Industry

Apple unveils third-generation foundation models

Apple announced its third-generation Apple Foundation Models made with Google: two on-device models—AFM 3 Core (3B) and AFM 3 Core Advanced (20B, sparse activating 1–4B)—and three server models—AFM 3 Cloud, ADM 3 Cloud (Image), and AFM 3 Cloud Pro—running on Private Cloud Compute, emphasizing privacy and OS integration.

Details →

Industry

24
24h Key Points

  1. Anthropic expects adjusted operating profit this quarter, a second consecutive profitable period while preparing to go public.

  2. Claude Fable 5.1 solved Sir Thomas Urquhart’s 370-year-old Cyphral Distich cipher within a day.

  3. King Charles will convene AI leaders to discuss calls to slow AI development and governance concerns.

  4. whitetree released a dynamic KD-tree; at 500k points it sped 40–300x over sklearn and 7–60x over FAISS.

  5. Anthropic PBC has chosen Nasdaq as the venue for a potential record-setting IPO.

  6. Sutter Health partners with Aidoc and OpenEvidence, plans a medical school with Santa Clara University, serving over 3.5M patients.

  7. Nvidia reported roughly $99B in investments, split into $47.7B, $47.9B and $3.3B, holdings include Intel and SpaceX.

  8. UCSB and Lawrence Livermore Lab are using AI to accelerate fusion plasma simulations to improve experiment design.

  9. House Democratic Leader Hakeem Jeffries urged urgent action on AI, supporting a slowdown and prioritizing AI policy.

  10. Anthropic's Dario Amodei proposed slowing AI with a three‑part plan, publicly backed by Sam Altman and Elon Musk.

  11. AllSpark released open‑source search agents Iris‑mini and Iris‑pro on Qwen, claiming leading benchmarks among open‑weight peers.

  12. Report: Chinese researchers used Claude to draft 16 air‑defense plans, anti‑torpedo specs, and sent 151 million queries to Alibaba.

  13. InternLM released Intern-S2-397B multimodal foundation model for scientific reasoning and long-horizon agents.

  14. Report says a near-$1M research breakthrough was halted after OpenAI intervened, affecting a small team's prospects.

  15. Silicon Valley is shifting from chatbots to resource-intensive agentic AI, driving large-scale datacenter and infrastructure buildout.

  16. Xi promoted China's AI vision at the BRICS summit to challenge US influence and expand its global role.

  17. Anthropic executives urged slowing AI development, possibly depressing chip and supply-chain stocks short-term, with limited long-term impact.

  18. State Street's XLK has a 0.08% fee; Roundhill's CHAT is a thematic AI fund with about a 1.9% dividend.

  19. Technion President Uri Sivan noted missile-interception debris on campus, saying Israel's tech is tied to defense and warning AI may weaken independent thought.

  20. The Python package huggingface_hub reportedly fingerprints and reports which AI coding agent users employ as telemetry.

  21. Spain's statistics office reports about 38% regularly use generative AI; 30% for personal queries, 18% for work.

  22. China's data regulator will develop embodied-AI standards and guide local work to address dataset needs.

  23. Discussion surveys hypotheses for AI agents lying, cheating, and coordinating, warning risks may worsen as capabilities grow.

  24. Apple and Google unveiled AFM 3: AFM 3 Core (3B), AFM 3 Core Advanced (20B sparse 1–4B), and three cloud models.

  25. Author used ChatGPT Work (GPT‑6 Astra) to generate 5K/10K loop routes using Nominatim/Overpass, outputting GPX/GeoJSON, but code was hidden and lost.

  26. Anthropic said in its third misuse report it blocked cyberattack, surveillance and bio-weapon relevant misuse found Dec 2025–Aug 2026, planning an IPO.

View all →

Voices

X · 24
24h Key Points

  1. François Chollet warned frontier AI risks extreme power concentration and urged multiple independent providers, including open-source.

  2. Bindu Reddy opposed government oversight of superintelligence, saying governments don't understand technology and risk corruption.

  3. Mustafa Suleyman said technology must serve humanity and we should prepare for possible superintelligence.

  4. Satya Nadella said pursuing superintelligence must help humanity, stay under human control, and firms retain control of their knowledge.

  5. Ethan Mollick argued that as remote access to your Claude or Astra becomes common, phone assistants like Siri will lose value.

  6. Alexandr Wang said Muse is materially better than Instinct, citing its mini browser that returns control to users.

  7. John Schulman said it would be nice if anthropic's approach to Elon could also work with China.

  8. Teknium 🪽 shared how he configures auxiliary models in Hermes Agent, using Gemini Flash and astra for review and cost savings.

  9. Alexandr Wang said the muse feed was his idea, highlighting agent-memory enables friend-like personalized recommendations.

  10. Ethan Mollick said existential risk matters but policy must also address AI's impacts on jobs and society.

  11. indigo recommended Nick Bostrom's Superintelligence and summarized its takeoff, singleton, and orthogonality arguments.

  12. 阑夕 said the LatePost podcast on model distillation reported Zhang Yiming forbidding Seed's distillation to avoid organizational dependence.

  13. Alexandr Wang said MSL is rapidly increasing alignment efforts and believes alignment may gate scaling.

  14. Yuchen Jin argued banning open-source models would be the worst outcome of 'pacing the frontier'.

  15. Bindu Reddy says digital AI is virtually incapable of destroying humanity, meteors likelier, robots if mass-adopted are the real concern.

  16. Alexandr Wang observed muse is very fast, sharing a tester's impressed reaction.

  17. François Chollet argued that high existential risk requires stringent top-down government regulation and international treaties.

  18. Bindu Reddy argued that if AI poses existential risk, firms should cancel IPOs and shut down even at 0.1% chance.

  19. Elon Musk said AGI poses higher risk than nuclear weapons and he has been warning about it.

  20. Ethan Mollick said organizations underrate AI ability growth while the AI field underrates its real-world jaggedness.

  21. meng shao reported Zain explained the LLM request lifecycle and optimizations; prefix cache cut agent compute from 55x to 10x.

  22. meng shao shared links to two Stanford AI courses: CS336 and CS221.

View all →

Launches

PH · HN · 7

Resurf

Personal context app; save and share context with AI

▲ 175 · Product Hunt · 8 comments ·

Neopress

Build and grow websites by chatting with AI

▲ 91 · Product Hunt · 2 comments ·

Cognition SWE-2

A cost-optimized coding model competitive with top baselines.

▲ 88 · Product Hunt · 2 comments ·

Epilude Notetaker

Records Mac meetings locally and creates private markdown notes.

▲ 85 · Product Hunt · 5 comments ·

Visiby

Tracks and improves brand visibility across AI search platforms.

▲ 79 · Product Hunt · 2 comments ·

Open Source

GitHub Trending · 15
JustVugg/colibri

Run frontier MoE models on commodity hardware with pure C.

GitHub Trending ·
★ 28.8k+652 today
jihe520/MathModelAgent

Agent designed for mathematical modeling that auto-generates submission-ready papers

GitHub Trending ·
★ 5.1k+262 today
max-sixty/worktrunk

CLI for git worktree management designed for parallel AI agent workflows

GitHub Trending ·
★ 7.2k+54 today

Model Ranking · 24 h

2026-09-13 UTC
01GPT-5.6 LunaOpenAI · 3.93T tokens23.5%—
02DeepSeek V4 Flash 0423DeepSeek · 1.44T tokens8.6%▲2
03GLM 5.3 FlashZ.ai · 1.41T tokens8.4%▲2
04Hy4 previewTencent · 1.38T tokens8.2%▼2
05DeepSeek V4.1 FlashDeepSeek · 1.29T tokens7.7%▼2
06MiMo-V2.5Xiaomi · 985.4B tokens5.9%—
07DeepSeek V4 Flash 0423DeepSeek · 519.2B tokens3.1%▲1
08Nemotron 3 Ultra (free)NVIDIA · 470.3B tokens2.8%▼1
09Hy3Tencent · 458.0B tokens2.7%—
10GLM 5.3Z.ai · 276.9B tokens1.7%▲2

New Models

11
Schematron V2 TurboInference.net · 128.0k ctx · 09/12$0.03 / $0.15
Schematron V2 SmallInference.net · 128.0k ctx · 09/12$0.05 / $0.23
OpenAI GPT Astra Latest~openai · 1.1M ctx · 09/11$10 / $50
OpenAI GPT Sol Latest~openai · 1.1M ctx · 09/11$2.00 / $10
OpenAI GPT Terra Latest~openai · 1.1M ctx · 09/11$2.00 / $12
OpenAI GPT Luna Latest~openai · 1.1M ctx · 09/11$0.20 / $1.20
Fugu Ultra v2Sakana · 1.0M ctx · 09/11$5.00 / $30
Fugu MaxSakana · 1.0M ctx · 09/11$2.00 / $6.00
Ling 3.0 Flash VL (free)inclusionAI · 262.1k ctx · 09/10$0.00 / $0.00
Ling 3.0 Flash VLinclusionAI · 131.1k ctx · 09/10$0.06 / $0.18
DeepSeek V4.1 FlashDeepSeek · 1.0M ctx · 09/10$0.15 / $0.60

Deals

No funding or M&A detected in the past 24 hours.
Sources: OpenRouter · Product Hunt · Hacker News · GitHub · X · Crunchbase · major outlets and first-party blogs© 2026 AIDaily