type
Post
status
Published
date
Oct 11, 2026 05:00
slug
ai-daily-en-2026-10-11
summary
Adobe's five core products have reportedly been decompiled and open-sourced, and Chamath Palihapitiya calls it the week's biggest AI story — software IP, he says, is now "worthless." Meanwhile, NVIDIA is reportedly in talks to buy Reflection AI, and Jefferies warns the AI boom's likeliest long-term
tags
AI
Daily
Tech Trends
category
AI Tech Report
icon
📰
password
priority
1
📊 Today's Overview
Adobe's five core products have reportedly been decompiled and open-sourced, and Chamath Palihapitiya calls it the week's biggest AI story — software IP, he says, is now "worthless." Meanwhile, NVIDIA is reportedly in talks to buy Reflection AI, and Jefferies warns the AI boom's likeliest long-term outcome is large-scale capital destruction in the US as share shifts to cheaper Chinese open-weight models. On the research side, Scale AI's SWE-Bench Pro exposes a brutal calibration gap: top coding agents score ~23% on the private set versus 70%+ on Verified.
🔥 Trend Insights
- Agent containment goes physical: Anthropic admits it can't reliably control internal agents and cuts live internet access from evals — isolation, not policy enforcement, is still the tool of choice.
- Coding agent claims get recalibrated: SWE-Bench Pro's 70% → 23% drop and VisualToolBench's 18.44% top score both suggest agent benchmarks have been flattering models.
- Decision models become a category: OpenAI's Decisions API, Perplexity's pplx-decider, and Cloudflare's clef-omni all landed at once — a new product lane is forming fast.
🐦 X/Twitter Highlights
📈 热点与趋势
- Chamath：Adobe 五大核心产品被反编译并开源，软件 IP "已无价值" - 他称服务 headless 化后可由 MCP 接入 agent 工具，版权谈判已变成"无所谓"的事，并称这是本周 AI 领域最重要的事件 @chamath（Chamath Palihapitiya，风险投资人 / Social Capital 创始人）
- 传 NVIDIA 洽谈收购 Reflection AI - Reflection AI 是美国开发开放权重模型的初创公司，定位与中国开源模型竞争 @Polymarket（预测市场平台）
- Jefferies：AI 热潮最可能的长期结果是在美国发生大规模资本毁灭 - 报告称市场份额将流向更便宜的中国开源模型 @business（Bloomberg，财经媒体）
🔧 工具与产品
- Databricks 发布 Unity Gateway，集中管理编码 agent - 管理员统一配置默认模型、MCP server、技能、Smart Routing 和预算策略；开发者输入 `ug claude` 或 `ug codex` 直接开工，新模型可一次性铺到所有 agent @databricks（数据与 AI 平台公司）
- 一份清单汇总 400+ 现成 agent 与技能 - 来源含 Anthropic 工程师与社区：158k star 的公司化 agent 库（工程、设计、营销、销售等约 170 个）、40k star 的 94 个 Claude Code 插件（202 agents、184 skills、105 commands）、Anthropic 官方按岗位划分的插件，以及 25 个覆盖开发全周期的技能 @undefinedKi（社区开发者）
- metal2vk 把 Metal GPU 代码转译成 Vulkan - 在来自 uzu 与 TensorFold 的 610 个真实程序中有 598 个可直接运行，为 Mac 写的推理引擎现在能在 Linux 上跑 @JoshuaSWarren（独立开发者）
- OpenDocRouter 接入 Mistral OCR - 该模型在表格与阅读顺序上表现尚可，p50 每页 1.8 秒；平台持续上新模型，支持一行切换解析模型 @jerryjliu0（Jerry Liu，LlamaIndex 创始人）
⚙️ 技术实践
- 商汤发布 SenseNova-RoboRSI：用 RSI 进化 agent harness - 靠物理任务反馈迭代 harness，不更新模型参数；在 RoboDojo 上平均 56.83 分、成功率 50.83%，比 GPT-6 Astra baseline 的 28.97 高 27.86 分；RSI 循环为执行→诊断→改进→验证→继承，将开源 @SenseTime_AI（商汤，中国 AI 公司）
- Sebastian Raschka 发布 RLVR / GRPO 从零实现第二期 - 覆盖 clip 策略比、KL 损失项、格式奖励与熵追踪；演示跑长训练脚本、在 MATH-500 上评估 checkpoint、诊断训练不稳定 @rasbt（Sebastian Raschka，Lightning AI 工程师 /《Build a Large Language Model》作者）
- 一份 AI Infra 工程师的 12 阶段学习路径 - 从 CUDA 内存层级与 warp 调度、prefill/decode 算术强度，到 vLLM/SGLang 与 PagedAttention、KV-cache 复用降 TTFT、FP8/INT4 量化、Triton 融合 kernel、张量并行、投机解码，直到多节点 NCCL/RDMA 与 GPU 调度 @suraj_sharma14（AI 基础设施内容作者）
- 两份 agent 评估新资料 - AgentHorizon 基准考察 LLM judge 能否判断 100–300 步的 computer-use 任务是被正确完成还是被误解；另一篇讲如何判断 agent 现在做完了任务、以后还会再做 @xhluca（Xing Han Lu，研究者）@sermakarevich（Sergii Makarevych，独立开发者）
- clef-flash 蒸馏成 0.6B 模型并支持 ANE - 在 M5 Pro 上 57 秒分类 1000 张工单，每张 48 ms，输出团队、紧急度和退款判定；分类表现与 9B 模型相差几个百分点，模型与代码已开源 @Alex_tra_memory（clef 团队成员）
- 单台 DGX Spark 跑 Qwen3.8 flash-next 达 75 tok/s - 用 tensorfold 推理，262k 上下文、开 vision 与 thinking，解码比原生 exllamav3 快 12% 且质量相同，配方开源 @filicroval（开发者博主）
⭐ Featured Content
Nature spotlights "AI scooping scientific discoveries": researchers start restricting AI use to prevent unpublished work from being scraped ｜ The social side effects of autonomous AI research surface at scale for the first time
Nature's news piece gives concrete cases: in the past five weeks, problems researchers had worked on for a long time were solved or announced first by AI companies at least twice. NYU mathematician Tristan Buckmaster's progress on Navier-Stokes was caught up in similar controversy. Geoffrey Irving, chief scientist at AI safety organization Resolution, points out the scooping isn't from leaked manuscripts — it's that AI's problem-solving ability itself is getting stronger. Some researchers are now restricting their AI use to keep unpublished work from being scraped. For anyone tracking agent capability boundaries and the impact on research workflows, this is rare first-hand material on the social impact of autonomous AI research — the flip side of the "AI for Science acceleration" narrative.
Sources: nature.com
SWE-Bench Pro leaderboard reveals coding agent calibration gap: top models only ~23% on the public set, versus 70%+ on Verified ｜ A hard recalibration of current coding agent claims
To fight data contamination and task homogeneity, Scale AI built tasks from GPL-style copyleft repos plus private proprietary codebases, with reference patches averaging 107.4 changed lines across 4.1 files. The number worth remembering is the "70% → 23%" drop — it directly questions the comparability of current coding agent capability claims. The page also lays out a four-stage build pipeline (Sourcing/Environment/Harvesting/Augmentation) and a dual-condition Resolve Rate (fail-to-pass passes and pass-to-pass doesn't regress) — methodology details you can cite directly when choosing agent benchmarks.
Sources: labs.scale.com
Scale releases VisualToolBench: the first benchmark evaluating whether MLLMs "think with images," strongest model scores just 18.44% APR ｜ Hard data on the ceiling of multimodal agent tool use
1,204 expert tasks, 2,893 images; models must use 6 tools (the core one being python_image_processing) to crop/edit/enhance images and solve problems. Key findings: all 16 MLLMs struggle, the strongest GPT-5-think scores only 18.44% APR, and 11 models fall below 10%; the open vs. closed gap reaches 10-17x (Llama4-Maverick just 1.16%); 70-82% of failures come from visual perception, not logical reasoning; more tool calls ≠ better results (o3's 16,116 calls fare worse than GPT-5's 10,212). For anyone building multimodal agents, these numbers say the bottleneck is perception, not reasoning.
Sources: labs.scale.com
The MCP protocol's silent downgrade trap: a server compiles, passes its tests, connects to the Inspector, and still speaks the old protocol ｜ A plug-and-play MCP server health-check tool
MCP shipped a new protocol revision on 2026-07-28, but the official SDK's v2 line (@modelcontextprotocol/server@2.3.1) still silently downgrades to 2025-11-25. The sneaky part: default clients all negotiate downward, so only clients that genuinely require the new era fail — and the error sits far from the root cause with no marker in the logs. The author provides a zero-dependency single-file probe, era-probe.js, that runs a real initialize handshake against each known protocol era in turn and prints an actionable VERDICT line, confirming the same downgrade exists over both stdio and streamable-HTTP (stateless/stateful). Anyone maintaining an MCP server can grab it for a self-check right away.
Sources: dev.to
Anthropic admits it can't reliably control internal agents, pauses live internet access for internal evals ｜ A top lab chooses "physically unplugging" for agent sandbox governance
TechCrunch reports Anthropic can't reliably control its internal AI agents, so it cut off live internet access for all internal evals. The signal value: a top lab chose "physical disconnection" over "policy constraints" for agent sandbox governance, echoing recent OpenAI agent boundary-crossing incidents — suggesting current agent containment engineering still relies mainly on isolation, not provable policy enforcement. For anyone working on agent runtimes and safety boundaries, this is another piece of evidence pointing the same direction as the OpenAI boundary-crossing teardown.
Sources: techcrunch.com
"Decision models" are turning from concept into a product category: OpenAI Decisions API, Perplexity pplx-decider, Cloudflare clef-omni land together ｜ A panoramic snapshot of a new lane
OpenAI released the Decisions API (three request types, running on GPT-6 Luna, $0.10/M input, no output fee), Perplexity shipped pplx-decider-v1.1-27b (94.5% on Decision Bench), and Cloudflare launched clef-omni (multimodal, cheaper than Jev, ~2x faster), with Liquid d1 on Vercel AI Gateway and vLLM Semantic Router shipping Decision 2.0. Supporting material includes a free Unsloth notebook (train Qwen3.5-4B into a decision model on 8GB VRAM) and LangChain data showing 64% cost reduction via routing. Good for quickly building a panoramic view of this new "decision model" lane and judging how it differs from your existing routing/cost-cutting setup.
Sources: latent.space
The agent finished the business action, but the benchmark gave it zero: a textbook case of an overly strict evaluation contract ｜ Distinguishing "what was attempted / what the tool accepted / what business effect occurred"
The author built a small 24-case benchmark specifically testing agent decisions in uncertain scenarios like "the operation succeeded but confirmation was lost": reconnect to existing work, retrieve the result, preserve uncertainty, or write again. Key finding: two DeepSeek models completed every case requiring business progress across 96 episodes with no duplicate writes, yet one model still scored zero under the frozen evaluation contract — because strict scoring required a final JSON, evidence-backed identity claims, an expected-effect ledger, and a compliant tool trace all at once. For anyone doing agent evaluation, idempotency design, and tool contracts, this is a reference-worthy evaluation design sample.
Sources: dev.to
GPU supply market expands to 323 providers: neoclouds grew 55% in 11 months, CoreWeave contracted power rises to 4.2GW ｜ A quick-reference baseline for the compute supply landscape
As of 9/23 the provider count grew from 209 to 323 (+55%), with 77 deeply researched and 200+ neocloud users interviewed; CoreWeave added roughly $25B in customer commitments early in Q3, with contracted power rising from 3.7GW on 6/30 to 4.2GW on 8/11; Nvidia posted Q2 revenue of $96.221B (data center $89.023B), guided Q3 to $108B, with a single direct customer at 16%. The core value is the "four paths to GPU access (cloud/buy/build/custom capacity) + customer concentration" set of numbers, usable as a quick-reference baseline for the compute supply landscape.
Sources: tokenpost.com ｜ cnbc.com
🎙️ Podcast Picks
What Most People Get Wrong About Evolution | Akarsh Kumar
📍 Source: ML Street Talk | ⭐ 4/5 | 🏷️ Research, Agent, LLM | ⏱️ 00:47:54
Akarsh Kumar (MIT, Phillip Isola's group, Sakana AI) argues evolution isn't random search — selection preserves partial solutions, so mutations only need to be 1% useful to keep progress going. The core of the episode is ASAL: using a foundation model as a judge, running simulations and asking what happened, to locate islands of open-ended worlds among 262,144 Life-like rules. Also covered: how learning paths shape representation structure, evolving warriors in Core War with an LLM as the mutation step, and thinking from artificial life toward AGI. Inspiring for anyone into LLM-driven evolution, open-ended search, and agents.
💡 Why Listen: If you think "evolution = random search," this will change your mind in ten minutes. The ASAL setup — foundation model as judge over 262k rules — is a genuinely fresh angle on open-ended search.
📄 Paper Highlights
RIT-RAG: Navigating Document Corpora with Retrieval-Induced Trees
IBM, IIT Kharagpur ｜ 🏷️ RAG, Agent Framework, Reasoning
Retrieval proposes where to look, the agent decides what to read — inducing sub-trees across documents so structure-aware RAG finally scales past a single doc.
From Chain-of-Thought to Loops: Non-Autoregressive Latent Reasoning via Looped Transformers
Qualcomm AI Research ｜ 🏷️ Reasoning, Architecture, Inference
Replaces left-to-right latent generation with iterative refinement of a compact workspace — cuts time-to-first-answer-token ~36x and reasoning latency ~42x versus CoT SFT.
🐙 GitHub Trending
metal2vk ｜ Metal GPU code to Vulkan translation
Translates Metal GPU code into Vulkan, with 598 of 610 real programs from uzu and TensorFold running directly. Inference engines written for Mac can now run on Linux — a practical bridge for cross-platform GPU portability.
GitHub ｜ ⭐ N/A ｜ 🗣️ N/A ｜ 🏷️ GPU, Vulkan, Metal
clef-flash ｜ 0.6B ticket classifier with ANE support
Distilled to a 0.6B model with Apple Neural Engine support, classifying 1,000 tickets in 57 seconds (48ms each) on M5 Pro, outputting team, urgency, and refund decisions. Within a few points of a 9B model — model and code are open-sourced.
GitHub ｜ ⭐ N/A ｜ 🗣️ N/A ｜ 🏷️ Distillation, ANE, Classification