待翻译：[AINews] SpaceXAI Grok 4.6 and Grok @Bot

AI 服务暂时不可用，以下为来源正文，待恢复后补全翻译。

One of our top recurring themes of the year has been coding agents breaking containment into knowledge work, and it’s clear that the AI teammate/multiplayer/multiagent space is the next big AI battleground. With Claude Tag launching to mixed reviews and Block’s Buzz requiring a more technical user, the space was still open for a new category leader, which the now Cursor→SpaceX team has adroitly shipped to very positive reviews: This is powered by their newest model, Grok 4.6, released today as arguably the second best knowledge work model in the world (as both competitor Cognition and Elon acknowledges)… though it is surely the top by efficiency: Grok 4.6 is a confirmed 1.5T model that “builds on Grok 4.5 with a particular focus on long-running agents and more ambitious interactive and visual work”. The only training disclosure can be reproduced in full (emphasis ours): Grok 4.6 underwent a longer supplemental training run than Grok 4.5, with curated model-generated data for reasoning and advanced technical concepts, high-quality engineering data, and an improved optimizer and training recipe. This produced a stronger foundation for the SFT and RL stages that followed. We then used Grok 4.5 to regenerate the SFT trajectories across reasoning efforts, agent harnesses, and domains such as STEM, software engineering, and knowledge work, and filtered out problematic traces with model-based checks. The resulting SFT checkpoint shows strong performance and improved behavior. Grok 4.6 is trained on a wide range of agentic RL tasks, including knowledge work, general coding, and domain-specific environments for kernel optimization, web development, computer-aided design, and more. It is currently unclear if the lack of sandbox escape incidents when it came to training Grok 4.6 is a testament to their infra engineers or an indictment of their researchers. (that is a joke about current events, don’t get mad) AI News for 8/11/2026-8/12/2026. We checked 12 subreddits, 544 Twitters and no further Discords. AINews’ website lets you search all past issues. As a reminder, AINews is now a section of Latent Space. You can opt in/out of email frequencies! AI Twitter Recap Frontier Model Day: Grok 4.6, Qwen3.8-Max, DeepSeek V4 Pro, and Microsoft’s MAI-Thinking-1 Grok 4.6 reaches the frontier on price/performance: xAI released Grok 4.6, described as a major step up from 4.5 at the same price. Independent evaluations from Artificial Analysis place it at 61 on the Intelligence Index, roughly in line with GPT-5.6 Sol Max, behind Claude Opus/Fable, with strong agentic results including 88.4% on Terminal-Bench v2.1, 1753 GDPval-AA v2 Elo, and competitive AA-Briefcase performance at far lower cost (AA-Briefcase note). Early arena data from Code Arena also slots it near GPT-5.6 Sol and Claude Fable on webdev tasks. Pricing is a central theme: AA highlights $2/$6 per 1M input/output tokens, materially below frontier peers, while practitioners immediately framed it as the new default for coding and bug-finding workloads (Pawel Huryn, Cognition availability in Devin). xAI says the gains came from a longer supplemental training run, regenerated SFT traces, and agentic RL over coding, web, CAD, and kernel optimization; they also report more self-testing behavior during long tasks (@kimmonismus summary). Elon also said Grok 4.7 is already in flight, with initial training complete and supplemental training on SpaceX internal data planned. Qwen3.8-Max open weights are out: Alibaba’s Qwen3.8-Max dropped as an open-weight 2.4T total / 95B active MoE. Community notes emphasize its scale, day-0 serving, and long-context/agent orientation: Yuchen Jin called it one of the largest open-weight releases to date; vLLM shipped day-0 support plus vendor-specific 4-bit checkpoints for NVIDIA B300 and AMD MI355X; Together AI and Baseten also announced immediate support. One important caveat from users: the released open-weights variant appears to be text-only, with no vision input in the initial drop (skalskip92). DeepSeek V4 Pro GA undercuts the market: DeepSeek’s V4 Pro GA rollout immediately drew attention less for “best benchmark in every column” than for economics. Multiple observers highlighted pricing around $0.435/M input and $0.87/M output (kimmonismus), with Cline calling it roughly 57× cheaper than Fable 5 while reporting meaningful gains over the preview, including a 15.8% Terminal Bench increase. Reaction was mixed on capability: some early users found it solid but not clearly ahead of Kimi/Flash on all tasks (Yuchen Jin’s roundup, scaling01, teortaxesTex), suggesting DeepSeek’s next gains may depend more on RL environment and agent work than raw scale. Microsoft enters with its own reasoning model: Mustafa Suleyman announced MAI-Thinking-1, Microsoft’s first reasoning model “built from scratch,” now available in Foundry. The initial ask from the team is notably practical—Finbarr Timbers specifically requested feedback on tool use—which suggests Microsoft is positioning it as an applied reasoning model rather than just a benchmark entrant. Solar Pro 4 also moved up a tier: Artificial Analysis reported that Upstage’s Solar Pro 4 jumped from 14 to 42 on the Intelligence Index, with especially large gains on agentic and long-context tasks, though still behind the current top frontier and open leaders on both raw score and price. Open-Weight Multimodal and Edge Models: Video, Vision, Voice, and Local Inference LTX-2.5 and the open video stack keep improving: @RisingSayak highlighted that Lightricks’ LTX-2.5 landed in Diffusers with several practical features that matter for local workflows: joint video + 48 kHz audio generation, prompt-controlled clip length, a 2-pass quality mode, tile rendering for lower memory usage, and preprocessing that re-compresses input images to better match training. Ostris AI Toolkit added support the same day. More broadly, several accounts framed the week as an unusually strong run for open multimedia releases, including MiniMax H3, LTX-2.5, LFM2.5-VL-3B, and North Micro Vision (victormustar, multimodalart). Small VLMs and local multimodal are getting serious: Cohere launched North Micro Vision, an Apache-2.0 open-source small VLM aimed at document understanding, with claims of outperforming Gemma 4 E2B and Ministral 3 3B on a broad visual benchmark mix (results thread). Liquid AI’s LFM2.5-VL-3B was also repeatedly cited as a strong compact vision model, and users demonstrated hybrid local/remote agent stacks—for example, Hermes Agent using DeepSeek V4 Flash for planning plus LFM2.5-VL-3B for local vision. Speech and sign-language releases were unusually substantive: Google DeepMind announced SL2T, a sign-language-to-text system powering ASL input on Android/Pixel 11. The follow-up notes are technically interesting: body pose tracking happens on-device, translation runs server-side, and the system is optimized for real-world constraints like one-handed signing (detail). Separately, Deepgram launched Flux TTS, a low-latency conversational TTS model claiming ~80 ms response time and mid-call adaptation for voice agents. Inference, Compression, and Systems: vLLM, Quantization, CUDA Scheduling, and Ranking Infra vLLM added important infra for giant models and long prompts: vLLM now supports Azure Blob paths for both model loading and KV connectors. The Microsoft/NVIDIA recipe matters operationally: faster weight loading via Dynamo ModelExpress (up to 7.3× faster on H100/A100) and blob-backed KV caching via LMCache + NIXL, trading recomputation for fetches on long-prompt workloads (follow-up). Compression work is extending the useful life of very large models: LLM Compressor v0.13.0 added REAP expert pruning for MoE models—dropping whole experts based on calibration saliency before quantization—as well as arbitrary 3/5/6/7-bit quantization. On the more extreme end, Unsloth claimed to shrink Qwen3.8-2.4T-A95B from 4.9 TB to 397 GB via dynamic 1-bit quantization, making local execution conceivable on 410 GB+ RAM/VRAM systems. They also showed a 2-bit Nemotron 3.5 Lightning setup sustaining long tool-use sessions in 22 GB VRAM. GPU kernel authoring is getting safer and more declarative: maharshii highlighted CuTeDSL 4.7.0 Task Scheduling kernels, which let developers explicitly declare warp roles, resources, dependencies, and schedules, enabling static checks for deadlocks, races, and barrier initialization before lowering to GPU code. The same author also posted a concise explainer on the prerequisites behind TMA async copy—acquire/release semantics, mbarriers, and CuTe arithmetic tuples—for people trying to reason about modern NVIDIA memory movement primitives (thread). Classic recommender/ranking stacks are still quietly delivering wins: François Chollet pointed to Expedia’s migration to a modern Keras 3 setup, reporting 30% faster training and 70% lower inference latency for ranking models (tweet). His follow-up stresses a more strategic point: Keras’s backend-agnostic APIs reduce lock-in if teams later need PyTorch or JAX kernels (note). Agents, Harnesses, and Developer Tooling: Reliability, Memory, Plugins, and Security The stack above the model is becoming the main product surface: Several tweets converged on the same theme: many practical gains are coming from harness engineering, memory, approvals, evals, and tools more than bespoke model training. Scott Stevenson restated the argument that RAG and harness engineering beat training most of the time because they personalize per customer, avoid privacy risks, improve in real time, and inherit base-model progress (thread, follow-up). Random Walker added a useful product distinction between delegation agents and collaboration agents, with very different optimization targets around verifiability, latency, and human control (tweet). Tooling releases reflected that shift: GitHub’s @code introduced Agent Plugins 1.0, packaging skills, MCP servers, and AI extensions together, and separately shipped UX improvements like sticky scroll and better session handling (release thread). OpenAI/Codex-side momentum showed up too, including Codex for Linux. LangChain rebuilt LangSmith dashboards for more useful trace analysis and reporting. Memory and portable agent state are becoming baseline expectations: Hermes Agent got multiple ecosystem updates, from Raspberry Pi deployment to easy profile export/import and new skills like generating reusable APIs from observed web traffic (Teknium). Managed Deep Agents examples from LangChain focused explicitly on durable memory and recurring workflows such as social-media agents (hwchase17). Security and governance for agents is becoming concrete: W&B showed a side-by-side agent email example where one agent leaked SSN/card info while another blocked prompt injection and redacted secrets before the model saw them (thread start). The Turing Post raised a more architectural issue around delegated identity: if an agent uses your SaaS credentials directly, revocation and auditing become muddy (tweet). Benchmarks, Research Directions, and AI-for-Science AI-assisted math and science claims are getting harder to ignore: The most engaged technical tweet was Steven Strogatz sharing a story that a neurosurgery resident reportedly used ChatGPT 5.6 to solve a significant open problem in numerical linear algebra (tweet). Relatedly, multiple accounts noted another EpochAI open problem apparently falling (scaling01). New benchmarks target less gamed capabilities: Princeton/MIT collaborators released DiG-bench, a text-based benchmark for discovery rather than standard QA or code tasks; tri Dao specifically praised it for having some of ARC’s flavor without confounding vision issues (tweet). Redwood + Anthropic introduced the Conceptual Reasoning Index, targeting AI-risk-relevant argumentation and conceptual reasoning where feedback is sparse and hard to automat [truncated for AI cost control]