Ox Alpha·Grok 4.6 코딩 경쟁, Starlink 우주 인프라, 장기 작업 에이전트 평가

Elon Musk

@elonmusk

Grok 4.6 on extra high thinking mode now achieves #1 score on CursorBench!

Ox Alpha free model beating Fable-5, GPT-5.6-sol, Grok 4.6, and GLM-5.3 on coding tasks. - Ox Alpha dropped 80% on a hard coding suite a 10-task DeepSWE subset. - Most frontier models are sitting at 52–65%. - 1M context, Multimodal, Zero data retention & Near-unlimited usage. Fingerprinting points strongly toward a next-gen GLM, but nothing is confirmed. This might be the strongest free coding model available right now. It’s free for the next week on OpenCode & OpenRouter.

Elon Musk

@elonmusk

Starlink now has 11k satellites in orbit

Starcloud, a startup building data centers that run AI workloads in orbit, raised $250 million at a $2.3 billion valuation, up from $1.1 billion in March. NVIDIA and Cisco Investments joined the round, and Starcloud plans to fly its largest orbital data center spacecraft on SpaceX's Starship.

Claude

@claudeai

Claude Security scans now run on Claude Mythos 5, available today in public beta for all Claude Enterprise customers. Put our most capable security model to work on your codebase, no separate model access needed.

NVIDIA AI

@NVIDIAAI

As AI agents take on more complex work, where should security live? The harness guides what an agent tries to do. Infrastructure controls what it can actually do. Our AI safety and security teams share how they’re thinking about security across the agent stack.

The largest fully annotated open egocentric human dataset. Today we're open-sourcing EgoSuite-Open100K with @huggingface : 100,000 hours of egocentric human data, free for anyone to use. Physical AI has a scaling law of its own now, and its input is human data. That part is settled. What the field still needs is a shared foundation to build on. 15,000+ tasks. 15,000+ real scenes, from factory floors to retail backrooms. Hand pose, body pose and subtask-level semantics, with wrist camera coverage on part of the set. Licensed for commercial training, not research-only. We build the data infrastructure for Physical AI. Open100K is the first public layer of it. The first 10,000 hours are live now, with the rest rolling out in stages.

François Chollet

@fchollet

This is very nice work from NVIDIA. Like all high-performing approaches on ARC-AGI-3, it uses deep learning-guided on-the-fly synthesis of symbolic world models, i.e. navigating the world by generating programs to represent what you know. To be clear, like with several other recent claims, scoring 100% on the public demonstration set is not the same as "scoring 100% on the ARC-AGI-3 benchmark". It would be like saying you beat a videogame because you cleared the tutorial level.

NVIDIA said its AVO agent architecture reached a perfect 100 score on the ARC-AGI-3 public benchmark, clearing all 183 levels across 25 environments. ARC-AGI-3 tests long-horizon autonomy by dropping agents into turn-based environments with no instructions or stated goals.

Traditional software is predictable. AI is not.  What fundamental skills do developers need in order to build and deploy AI applications? Here's Andrew Ng's list: LLM foundations Grounding models with data Building agentic systems Evaluation-driven development Operating in production Machine learning foundations This is the second installment of the AI Engineering Skills Map. Read more and subcribe: https:// hubs.la/Q04tRNDQ0 #AI #MachineLearning #AIEngineering

Viv

@Vtrivedy10

launching another update to this skill soon - open holy grail question if anyone wants to riff! what goes in “How-to-make-tasks-harder[.]md”? some options in a list below part of making a good eval is making sure it captures the real world. another part is making the actual task hard so there’s something to hill climb one way to calibrate “hard” is running a weaker model and smarter model, if everything passes all the time that’s not great -> perfect pass @ k isn’t a good learning signal a list of things that can make tasks harder, would love to compile more and hear other strategies - add more data (not super bullish on this as models are great brute forcers) - add information that needs to be discovered by search (ex: look across multiple tables to find missing information) - completeness (add many cases that all need to be passed, this works well. it can incentivize bad behavior though like over checking) - cross-domain tasks that combine 2 abilities. this is the best imo, agents are not good at this but it can look artificial and toyish humans are still useful here but would love to chat on thoughts

Banger paper from Google DeepMind. It's on the very hot topic of model routing. Routers assume the value estimate is free. Working out which specialist handles a query best costs money too, and it is often the larger cost in the pipeline. This new work formalizes model routing as Pandora's Box, the classic problem of optimal search when inspection is expensive. Under a Gaussian signal model the policies come out in closed form, telling you per specialist and per input whether refining the estimate is worth the price. Across a multi-LLM benchmark, retrieval-augmented specialists, and LLMs with variable inference-time reasoning, Pandora's Router matches exhaustive estimation quality while calling the expensive estimator far less often. When competing estimates are noisy, value-of-information reasoning raises the strategic specialist's utility at everyone else's expense. Paper: https:// arxiv.org/abs/2608.20316 Track more trending AI papers in our academy: https:// academy.dair.ai

Today we are open sourcing Ling-3.0-flash-dspark, a DSpark draft model built specifically for Ling-3.0-flash. On 4 NVIDIA Blackwell GPUs at batch 1, it delivered 1,120 tok/s, 0.78 ms mean TPOT, and an accept length of 9.95 across 1,000 requests.