1Claude Cowork Ships Live Artifacts That Stay Connected to Your Data

Claude just turned artifacts from static outputs into living dashboards. In Cowork, you can now build trackers and dashboards that connect directly to your apps and files — open one next week and it refreshes with current data instead of showing a stale snapshot. This is Claude graduating from "generate a thing" to "maintain a thing," and it changes the calculus for anyone who's been building internal tools by hand. (7,389 likes | 457 RTs) Read more →

2Kimi K2.6: A Trillion-Parameter Open Model Claims Opus-Level Coding

Moonshot AI's Kimi K2.6 packs 1 trillion parameters into an open-weight MoE model and claims Opus 4.6-level coding performance at 76% lower cost. If those benchmarks hold under independent evaluation, this is the most capable open model released this year — and it drops the same week as Qwen3.6-Max, making it an exceptionally competitive moment for frontier-class alternatives. (368 likes | 423 downloads) Read more →

Alibaba's Qwen3.6-Max-Preview Enters the Closed-Model Fight. After shipping open-weight Qwen3.6-35B last week, Alibaba now drops their strongest closed model yet. Qwen3.6-Max-Preview plays the same dual strategy as Meta — give away weights to build ecosystem, sell the premium tier to fund it. (509 likes | 259 RTs) Read more →

Google Opens AI Studio's Full Model Lineup to Paid Subscribers. AI Pro and Ultra subscribers now get increased limits and Nano Banana Pro access directly in AI Studio — no API key required. The barrier just dropped from "sign up for developer access" to "log in and go." (515 likes | 56 RTs) Read more →

Ollama Puts Kimi K2.6 on Cloud So You Don't Download 1TB: Same-day cloud integration means you can test a frontier-class trillion-parameter open model without pulling a terabyte of weights to your machine. Try it with Claude Code, your agent framework of choice, or just kick the tires from a browser. (227 likes | 31 RTs) Read more →

HuggingFace Opens 1M Spaces as Callable Agent Tools: Every one of HuggingFace's million Spaces is now invocable as a function by AI agents — turning HF into the largest tool registry for autonomous systems. If you're building agent pipelines, this is your new tool catalog. (866 likes | 109 RTs) Read more →

Claude Code v2.1.116: Resume Sessions 67% Faster: MCP startup is now parallelized, large session resume is up to 67% faster, and VS Code terminal scrolling is smoother. Quality-of-life improvements that compound for daily users — update now. Read more →

Inside Live Artifacts: How Connectors Power Claude's Dashboards: Felix Rieseberg's developer walkthrough explains how Connectors wire live artifacts to your data sources and which use cases actually work today. The practical companion to the official announcement — start here if you want to build your first live dashboard. (447 likes | 14 RTs) Read more →

3Opus 4.7 Quietly Costs 1.46x More Tokens Than 4.6

Simon Willison's token counter reveals the hidden cost of upgrading: Opus 4.7 consumes 1.46x more tokens for text and up to 3x more for images compared to Opus 4.6 — same per-token price, significantly higher bills. If you're planning a migration, audit your actual token usage on representative workloads before flipping the switch. Budget impact could be substantial for image-heavy pipelines. (1,510 likes | 132 RTs) Read more →

The Official Codex Beginner's Guide Treats It as a Platform: OpenAI's official RT of a comprehensive beginner's guide signals they're positioning Codex as a full platform, not just a CLI. Covers install, prompting, and agent configuration from zero — useful even if you've been using it casually. (2,272 likes | 257 RTs) Read more →

GGUF, Unsloth, Q_4_M, IQ_4XL: The Quantization Format Zoo Explained: HuggingFace finally published the clear explainer everyone browsing model pages has been needing — what each quantization format actually means, what you lose at each compression level, and when to pick which. Bookmark this before downloading Kimi K2.6's GGUF variants. (1,322 likes | 146 RTs) Read more →

Why a Single Benchmark Number Lies About the Open-Closed Gap: Interconnects breaks down why comparing open vs. closed models with one leaderboard score is fundamentally misleading — the gap shifts dramatically by task type, quantization level, and evaluation methodology. Essential context before you take Kimi K2.6's "Opus-level" claims at face value. Read more →

Noetik Aims Transformers at Cancer's 95% Trial Failure Rate: Ron Alfa and Daniel Bear are applying autoregressive transformers to clinical trial matching — reframing the 95% cancer trial failure rate as a sequence-matching problem rather than a biology problem. A compelling Latent Space deep-dive on transformers doing real work beyond text generation. Read more →

4Anthropic Secures 5 Gigawatts of Amazon Compute

Anthropic just locked in up to 5 gigawatts of compute capacity through Amazon, with nearly 1 GW expected by end of quarter. For context, 5 GW could power a mid-sized city. This isn't about training the next Claude — it's about having enough headroom to train several next Claudes simultaneously while serving inference at scale. The infrastructure arms race is now measured in power plant equivalents. (6,472 likes | 470 RTs) Read more →

LeCun Doubles Down: Generative AI Is a Dead End. Yann LeCun's long-standing critique of autoregressive generation gets another public airing — and on a day when a trillion-parameter autoregressive model just dropped, the timing is either terrible or perfectly illustrative depending on where you sit. His argument for fundamentally different architectures remains the most substantive counter-narrative to the scaling consensus. (3,793 likes | 651 RTs) Read more →

Anthropic Recruits Domain Scientists With STEM Fellows Program. The new program embeds researchers from science and engineering directly alongside AI teams — a signal that the next wave of Claude improvements will be driven by field-specific expertise, not just compute and data. (1,273 likes | 126 RTs) Read more →

Atlassian Quietly Turns On Default Data Collection for AI Training. Every Jira ticket and Confluence page in your org is now training data unless you've opted out. Atlassian flipped the default with minimal fanfare — if you're running enterprise Atlassian, check your data sharing settings today, not tomorrow. (476 likes | 111 RTs) Read more →

Chinese Tech Workers Push Back on Training Their AI Replacements. Companies are asking employees to create AI replicas of themselves — and workers are resisting. The "Colleague Skill" GitHub project that sparked this is a preview of workforce automation tensions coming to every tech company within the next two years. Read more →

Claude Code Hackathon Winners Show What Agent-Native Dev Looks Like: The "Built with Opus 4.6" hackathon winners are out, and the winning projects are concrete proof points for agent-native development patterns — not demos, but working tools that ship. Study the architectures if you're building with Claude Code subagents. Read more →

Qwopus-GLM-18B: A Community Cross-Architecture Model Merge: A community merge of Qwen and GLM architectures into a single 18B GGUF is trending on HuggingFace with nearly 3K downloads. The open-source community is now speed-running cross-architecture merging — test whether it outperforms either base model for your use case. (124 likes | 3.0K downloads) Read more →

Mixture of Experts (MoE): Kimi K2.6 packs 1 trillion parameters into an open-weight model that runs affordably — and the reason is MoE architecture. Instead of activating every parameter for every token (like a dense model), MoE routes each token to a small subset of specialized "expert" subnetworks. So while the total model is 1T parameters, any given inference might only activate 50-100B of them, keeping compute costs comparable to much smaller dense models. This is why "1T parameters" doesn't mean "1T cost" — and why MoE is becoming the default architecture for frontier-scale open models that need to be actually runnable.

• HuggingFace Opens a Tokyo Office: Serious investment in Japan's underserved open-source AI community. (2,100 likes | 301 RTs) Link
• Deezer: 44% of Daily Uploads Are Now AI-Generated: The first hard number from a major platform quantifying the AI content flood. (271 likes | 260 RTs) Link
• GitHub's Fake Star Economy Exposed: An investigation into star manipulation — cross-reference with download metrics before trusting counts. (729 likes | 354 RTs) Link
• Recursive Superintelligence Raises $500M at $4B — Four Months Old: GV-led, pre-product, purely on team and thesis. Peak AI capital gravity. (41 likes) Link

LeCun's dead-end thesis meets its trillion-parameter stress test. On the same day Moonshot ships a trillion-parameter autoregressive model claiming frontier-level performance, Yann LeCun once again declares that the entire autoregressive paradigm is a dead end. This isn't just academic contrarianism — it's the most coherent alternative thesis to the scaling consensus, and today's releases are the exact evidence both sides will cite to prove their case. The scalers will point to Kimi K2.6 and say "look, it keeps working." LeCun will point to the same model and say "you needed a trillion parameters and MoE tricks to match what a fundamentally better architecture could achieve at a fraction of the cost." What makes his position worth engaging with seriously is that he's not arguing against capability — he's arguing against efficiency. And in a world where Anthropic just secured 5 GW of compute and trillion-parameter models require Ollama cloud just to be testable, the efficiency question isn't academic anymore. Whether or not LeCun is right, his critique is the sharpest lens we have for asking: is this the best way to spend a civilization-scale investment in compute? Read more →

Until next time ✌️