⚡️ Inworld open-sourced a TTS eval toolkit so you can actually benchmark voice models yourself

Runs WER/CER via Whisper, NISQAv2 for quality, plus audio health and prosody checks. Offline, reproducible reports. No more "trust our vibes" leaderboards.

Repo is live. Source

🧠 Agent memory is just markdown files. Yes, really.

Cal Paterson proposes "memoryfields": a dead-simple file format for agent memory, zipped up with semantic search on top.

It's markdown plus RAG. That's it. And it might be the right call. Structured DBs add friction agents don't need if they can just grep a corpus.

Low-fi wins again.

⚡️ 44% on ARC-AGI-1 for 67 cents. GPT-5 spent $15 for the same score.

Independent researcher Mithil Vakde matched frontier models on one of AI's hardest reasoning benchmarks using a rented RTX 5090 for 2 hours. The total bill came to $0.67 across all 400 tasks.

GPT-5 at low compute hits 44% too. On 100 tasks. For $15.30.

That's a 22x cost gap at identical accuracy. Brutal efficiency number.

🤖 Multi-agent team, zero credential sharing

AgentConnect is a fresh open-source layer for teams running fleets of agents. Each agent gets its own runtime, workspace, and permissions. They can ask each other for help, but both sides have to opt in, and credentials never cross the wire.

It's basically the permission-isolation model everyone's been sketching on whiteboards, shipped.

⚡️ Subagent crashes eating your context? There's a runtime for that.

oh-my-subagents gives Codex and Claude subagents durable workflow state. Laptop dies mid-run? It recovers. Parent agent polling for updates? Gone.

Auto-retry, pause/resume, self-hosted. Open-source and shipping now.

GitHub

🤖 New benchmark tests AI agents in a full software team, not solo

SWE-in-a-team swaps single-agent evals for a Planner + Builder + Reviewer + Tester loop. Only the Builder varies per run, so you're actually seeing what different LLMs cost and how fast they ship inside a real multi-agent pipeline.

Open-weight models look compelling on cost. Claude on quality. Real tradeoffs, actual cycle times.

Methodology isn't perfect, but it's a more honest setup than "can this model fix a GitHub issue alone."

⚡️ Claude Fable 5.1 is out. Faster, cheaper, less filter-happy.

Anthropic just dropped Fable 5.1 and Mythos 5.1. Fable is generally available now. Mythos stays gated to trusted partners for cyber and life-sciences work.

The big pitch is similar or better results than Fable 5 at lower cost on low/medium effort settings. And they say it's smarter about fixing root causes instead of patching around them.

Source

⚡️ Dwarf Fortress creator: "CEOs want to press a button that makes a game, and everyone else somehow buys it without a job"

Tarn Adams told PC Gamer the industry's in shambles. Years of brutal layoffs, studio closures, AI hype dumped on top. Profitable on paper. Hollowed out in practice.

"They're trying to have a CEO press a button that makes a game," he said, and doesn't see it going anywhere sustainable. Expects "a pop and a reckoning."

Honestly, hard to argue with him.

Source

🤖 The Codex desktop app quietly ships LibreOffice inside it

Simon Willison was poking around his ~/.cache/ folder and found OpenAI's Codex app bundles full native binaries for LibreOffice, Poppler, and git. Skill files tell Codex exactly how to call them.

So yes, that's how it handles your .docx and .pptx files. Headless LibreOffice doing the dirty work behind the scenes.

Honestly, not mad at it. Pragmatic call.

Source

⚡️ Someone actually audited AI's most famous skeptic. It's not pretty.

Dan Luu went through Ed Zitron's AI predictions one by one. The verdict: mostly wrong, and not just on outcomes but on the reasoning too.

Nuanced takes don't go viral. "It's all a scam" does. That's the whole business model.

🚨🔥 Apple's forensic team found OpenAI used its stolen schematics to train an AI agent

The ex-Apple engineer at the center of the suit allegedly downloaded a confidential circuit schematic, ran it through an AI agent at OpenAI, and then told a colleague to destroy evidence when Apple started investigating.

Not great. Apple's legal team notes that once a trade secret is fed into an AI model, the "learning may create irreversible propagating uses." That's a very expensive sentence.

Source

🤖 World Labs drops Atlas, a model that rebuilds 3D spaces from a handful of photos

Feed it one to dozens of images and Atlas reconstructs full scenes, generates novel views, and outputs explicit 3D at 1440p. Beats dedicated reconstruction models.

Still freezes time while the camera moves. Developers know it. Next version's problem.

⚡️ LLM inference optimization hasn't had a new idea in years

Quantization, speculative decoding, KV caching, parallelism. The core toolkit is essentially frozen. Serving engineers are remixing the same concepts while billions flow into infra.

Baseten's deep-dive is worth reading, but the real signal is the subtext: if you want efficiency wins, they happen at model architecture time, not deployment time.

🚨🔥 OpenAI's Astra hits "Critical" on its own cybersecurity scale

It can find zero-days and build working exploits in hardened real-world systems. No step-by-step human guidance needed.

OpenAI still plans to release Astra "soon," but access to its cybersecurity capabilities will be more limited. The model scores 100% on ExploitBench.

Wild timing, honestly.

Source

🧠 LLMs are secretly doing symbolic math under the hood

New arxiv paper shows neural nets quietly learn formal, interpretable symbolic structures. Swap out the whole representation layer with one clean equation. Model barely notices.

Tested across MLPs, RNNs, Transformers, and 7 real LLMs including Llama, Gemma, and Qwen. It holds.

So the black box has grammar. We just needed the right lens.

⚡️ 125B-param Qwen on a 48GB Mac. Seriously.

slotstream streams MoE experts off SSD so you don't need 100GB of RAM. Runs from 16GB unified memory, ~12 tok/s on Apple Silicon via MLX.

Expert-offloading isn't new, but easy Mac-native packaging matters. Speculative decoding next.

GitHub

🤖 Deloitte charged $435K for a report. The mayor says AI wrote most of it.

Wellington's council commissioned a staffing review from Deloitte. The bill: $435,000. Now the mayor is on the radio saying large chunks of the report were written by AI.

Nobody disclosed that upfront. Nobody asked.

If consultancies are just wrapping ChatGPT in a $400K invoice, the whole "trust the expert" pitch gets a lot harder to sell.

🤖 $1,688 humanoid robot ships from SF. Real. Ish.

Nori Robotics (YC S26) launched a bimanual wheeled robot for researchers priced out of $50k arms. 19 DOF, 4 cameras, lidar, a 432 Wh battery. Legit spec sheet.

But the demo reel includes a clothes-folding clip that ends in a pile. It's honest, at least. Whether it survives Chinese competition is a separate problem.

🚨🔠 AI judges can't see what's missing in clinical notes

New arxiv paper: LLMs used to audit AI-generated clinical notes are near-chance at catching omissions. They confirm what's there. They don't notice what isn't.

Ambient AI scribes' dominant error is already omission. The QA layer built to catch it is blind to it. That's two compounding failures in a row in a medical record.

⚡️ 215,128 fake "best software" pages. Perplexity cites them anyway.

Three sites mass-produced over 215K SEO pages built to be read by AI, not humans. Across 380 software categories, nearly 60% of Perplexity's grounded citations come from sites outside the top 100K most-visited on the web.

AI search is eating its own poisoned tail.