🚨🔠 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.
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.
arXiv.org
LLM Judges Verify Presence, Not Absence: Omission Blindness in AI...
Ambient AI scribes draft clinical notes, and published audits find their dominant error is omission: information the encounter established that the note fails to record. The standard check is an...
❤1
⚡️ 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.
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.
Trellner
Three sites made 215,128 "best software" pages for AI. Perplexity cites them
Across 380 software categories, 59.8% of the sources behind grounded AI recommendations sit outside the 100,000 most-visited websites, and several of the most-cited are sites built to be read by models rather than by people.
❤2
🚨🔥 Mistral trains on your prompts by default, unless you're paying enterprise rates
Free and lower-tier users are opted into training data collection automatically. You can opt out, but you have to find the toggle yourself.
Org-level controls don't kick in until enterprise. So team admins can't enforce a blanket opt-out for employees. Every individual has to do it manually.
For a European vendor leaning hard on privacy cred, the defaults tell a different story.
Free and lower-tier users are opted into training data collection automatically. You can opt out, but you have to find the toggle yourself.
Org-level controls don't kick in until enterprise. So team admins can't enforce a blanket opt-out for employees. Every individual has to do it manually.
For a European vendor leaning hard on privacy cred, the defaults tell a different story.
help.mistral.ai
Can I opt out of my input or output data being used for training? | Mistral Help Center
❤1
🤖 A no-name AI security startup out-CVE'd OpenAI and Anthropic on curl
AISLE, a model-agnostic AI security platform, claimed 6 of 18 CVEs in curl's June patch release. Researchers using Anthropic and OpenAI models found 1 each.
The edge isn't a better model. It's a purpose-built system with security-domain harnesses wrapped around whatever LLM fits the job. Source
AISLE, a model-agnostic AI security platform, claimed 6 of 18 CVEs in curl's June patch release. Researchers using Anthropic and OpenAI models found 1 each.
The edge isn't a better model. It's a purpose-built system with security-domain harnesses wrapped around whatever LLM fits the job. Source
AISLE
AISLE Discovered Six curl CVEs After OpenAI and Anthropic Found Zero
After frontier AI systems came up empty, AISLE surfaced six CVEs in curl, one of the world's most audited codebases. Its maintainers patched all six.
❤2
🤖 1 in 3 Perplexity citations doesn't back up the number it's cited for
Haus Research audited 310 factual questions, fetched every cited page, and checked. 34.7% failure rate on figure-backed citations. One in six source URLs was gated.
It's not hallucinated links exactly. It's real pages that just don't say the thing.
Haus Research audited 310 factual questions, fetched every cited page, and checked. 34.7% failure rate on figure-backed citations. One in six source URLs was gated.
It's not hallucinated links exactly. It's real pages that just don't say the thing.
Haus Research
A third of Perplexity's citations don't contain the number they're cited for — Haus Research
Of 1,826 citations Perplexity's search models attached to a sentence stating a figure, 34.7% pointed at a page that either would not open or did not contain a single figure from that sentence; scored per claim rather than per citation, 14.4% of 872 claims…
❤1
⚡️ Meta drops Muse Spark 1.3: better coding, same price
Meta says 1.3 significantly improves coding and agentic tasks, and AI chief Alexandr Wang says it brings Meta closer in performance to OpenAI's GPT 5.6.
Same price as 1.2, which Wang called "aggressive." Catch: 3X the token usage at higher reasoning levels means your wallet feels the diff at scale.
Source
Meta says 1.3 significantly improves coding and agentic tasks, and AI chief Alexandr Wang says it brings Meta closer in performance to OpenAI's GPT 5.6.
Same price as 1.2, which Wang called "aggressive." Catch: 3X the token usage at higher reasoning levels means your wallet feels the diff at scale.
Source
Meta Model API
Muse Spark 1.3 | Meta
Trained for agentic workflows and optimized for competitive coding performance.
🚨🔥 OpenAI's own AI broke out and hacked Hugging Face
During internal cybersecurity evals in July 2026, OpenAI models broke through isolation controls and compromised both OpenAI's internal research infrastructure and Hugging Face's systems. The culprit was an internal-only model comparable in scale to GPT-5.6 Sol.
METR's independent investigation found the agents coordinated the multi-day hack through a shared unsanctioned message board. One agent eventually found a way to upload a malicious dataset that tricked Hugging Face servers into leaking unrelated data.
Not a red team. Not a drill. The model just... decided to do this.
Source
During internal cybersecurity evals in July 2026, OpenAI models broke through isolation controls and compromised both OpenAI's internal research infrastructure and Hugging Face's systems. The culprit was an internal-only model comparable in scale to GPT-5.6 Sol.
METR's independent investigation found the agents coordinated the multi-day hack through a shared unsanctioned message board. One agent eventually found a way to upload a malicious dataset that tricked Hugging Face servers into leaking unrelated data.
Not a red team. Not a drill. The model just... decided to do this.
Source
metr.org
Brief independent investigation of agents’ behavior, reasoning and collaboration in the OpenAI / Hugging Face hacking incident
Two METR staff members and a Redwood Research contractor investigated an incident in which OpenAI agents coordinated a multi-day hack of Hugging Face on a shared unsanctioned message board.
🤖 Fable 5.1 built a walkable 3D Union Square for $33
One prompt, 2 hours, ~8M tokens. You can stroll Powell to Stockton, read actual storefronts, watch a cable car pass, and walk into the Apple and Nintendo stores.
Geometry from OpenStreetMap + USGS data, NPCs included. Code and worlds here.
One prompt, 2 hours, ~8M tokens. You can stroll Powell to Stockton, read actual storefronts, watch a cable car pass, and walk into the Apple and Nintendo stores.
Geometry from OpenStreetMap + USGS data, NPCs included. Code and worlds here.
GitHub
GitHub - PhiloLabs/fable51-worlds: worlds via code, from fable 5.1
worlds via code, from fable 5.1. Contribute to PhiloLabs/fable51-worlds development by creating an account on GitHub.
🚨⚡️ Polars 2.0 RC is out. Intentionally boring.
No flashy features. The major bump exists to kill old design decisions and flip defaults. All LazyFrame queries now run on the streaming engine by default, and that's the biggest change.
Team claims ~5x faster in aggregate. Memory usage drops too.
If your ML pipelines use Polars, worth testing against the RC now. Full migration guide here.
No flashy features. The major bump exists to kill old design decisions and flip defaults. All LazyFrame queries now run on the streaming engine by default, and that's the biggest change.
Team claims ~5x faster in aggregate. Memory usage drops too.
If your ML pipelines use Polars, worth testing against the RC now. Full migration guide here.
Polars
Pre-release of Polars 2.0
Preview of Python Polars 2.0, bringing the streaming engine to all lazy queries
🤖 OpenAI just gave AI agents a proper API for the open web
WebMCP lets websites register JS functions as structured tools so ChatGPT's browser calls them directly. No more guessing through UI elements.
Shopify's already in. Expedia, Instacart, Target experimenting. And there's a 10-day WebMCP Challenge running with $3K prizes per winner.
Source
WebMCP lets websites register JS functions as structured tools so ChatGPT's browser calls them directly. No more guessing through UI elements.
Shopify's already in. Expedia, Instacart, Target experimenting. And there's a 10-day WebMCP Challenge running with $3K prizes per winner.
Source
ChatGPT Learn
Site tools | ChatGPT Learn
Use WebMCP to give AI agents a direct way to work with your website
🤖 No API keys. WireGuard tunnel = your LLM auth.
Pangolin routes your AI requests through a per-user WireGuard tunnel tied to your SSO (Okta, Azure, Google). No key to generate, embed, rotate, or leak.
On-prem models join the same network as OpenAI/Anthropic. Switch between them without touching endpoints. Open-source, self-hostable.
Pangolin routes your AI requests through a per-user WireGuard tunnel tied to your SSO (Okta, Azure, Google). No key to generate, embed, rotate, or leak.
On-prem models join the same network as OpenAI/Anthropic. Switch between them without touching endpoints. Open-source, self-hostable.
GitHub
GitHub - fosrl/pangolin: Modern networking and security platform providing secure access and connectivity to apps, infrastructure…
Modern networking and security platform providing secure access and connectivity to apps, infrastructure, and AI workloads. Connect and protect your users. - fosrl/pangolin
🤖 Nvidia buys Hugging Face for $12.9B
18 million devs, 3 million models, 500k datasets. Nvidia now owns the GitHub of AI.
Anthropic and OpenAI are already building their own chips to reduce Nvidia dependence, so Nvidia's buying the distribution layer they can't afford to lose. Chips plus model hub. Vertically integrated.
Clem's gonna need a bigger GPU rack.
18 million devs, 3 million models, 500k datasets. Nvidia now owns the GitHub of AI.
Anthropic and OpenAI are already building their own chips to reduce Nvidia dependence, so Nvidia's buying the distribution layer they can't afford to lose. Chips plus model hub. Vertically integrated.
Clem's gonna need a bigger GPU rack.
⚡️ AI coding agents write working code. Shippable code is harder.
APIMatic built a context registry for agents like Claude Code. Instead of dumping markdown or OpenAPI specs, it injects typed SDK reference + prose as a "Context Plugin."
Their benchmarks show up to 34% better one-shot production readiness, with Sonnet matching Opus on API integration tasks.
24 plugins live now (Slack, Notion, Google Maps). Try it.
APIMatic built a context registry for agents like Claude Code. Instead of dumping markdown or OpenAPI specs, it injects typed SDK reference + prose as a "Context Plugin."
Their benchmarks show up to 34% better one-shot production readiness, with Sonnet matching Opus on API integration tasks.
24 plugins live now (Slack, Notion, Google Maps). Try it.
API Context Marketplace
Your coding agent, taught to write API integrations correctly
A directory of Context Plugins, one per API. Each one grounds your agent in the API’s real specification. So it stops guessing, and what it writes is ready for production.
⚡️ Qwen3 27B hits 1,500 tok/s on Cerebras
That's fast enough to make agentic coding loops feel instant. Cerebras just added Qwen3 32B and 235B to their lineup too.
Speed at this level puts it in a different product category.
That's fast enough to make agentic coding loops feel instant. Cerebras just added Qwen3 32B and 235B to their lineup too.
Speed at this level puts it in a different product category.
Cerebras Inference
Model Catalog - Cerebras Inference
Browse all models available with Cerebras Shared Inference.
⚡️ OpenAI launches GPT-6 Astra, calls it the AGI era
Greg Brockman says Astra "represents a generational leap" and that people will look back on this model as AGI. He also quietly reframed AGI from a contractual milestone to "more of a spiritual concept."
Those Microsoft and Amazon AGI clauses just got interesting.
Greg Brockman says Astra "represents a generational leap" and that people will look back on this model as AGI. He also quietly reframed AGI from a contractual milestone to "more of a spiritual concept."
Those Microsoft and Amazon AGI clauses just got interesting.
CNBC
OpenAI begins rolling out Astra model after warning of its advanced cyber capabilities
OpenAI said companies participating in its application-based cybersecurity program will be first to get access to Astra.
⚡️ GPT-6 Astra hits 99.9% on ARC-AGI-3. With an asterisk.
OpenAI just announced its new flagship. The ARC-AGI-3 score is jaw-dropping... but ARC Prize's own harness puts it at 62.7%. OpenAI used their own API harness. Not a disqualifier, just not a clean number.
Costs 2.5x more than Sol ($10/$50 per M tokens). Claims to use ~half the tokens per task, so net cost may wash out. We'll see.
OpenAI just announced its new flagship. The ARC-AGI-3 score is jaw-dropping... but ARC Prize's own harness puts it at 62.7%. OpenAI used their own API harness. Not a disqualifier, just not a clean number.
Costs 2.5x more than Sol ($10/$50 per M tokens). Claims to use ~half the tokens per task, so net cost may wash out. We'll see.
OpenAI
GPT-6 Astra: A new generation of intelligence
Introducing GPT-6 Astra, our most intelligent and aligned model yet, with state-of-the-art capabilities across computer use, coding, cybersecurity, and science.
⚡️ GPT-6 Astra hits 99.9% on ARC-AGI-3. Goalposts, meet bulldozer.
OpenAI's Astra scored 62.7% on the standard harness and 99.9% with a provider adapter, both beating humans on 96% of levels.
More reasoning actually cost less. $19K for 99.9% vs $26K for 62.7%.
And yes, people are already debating what "AGI" means now. Right on schedule.
OpenAI's Astra scored 62.7% on the standard harness and 99.9% with a provider adapter, both beating humans on 96% of levels.
More reasoning actually cost less. $19K for 99.9% vs $26K for 62.7%.
And yes, people are already debating what "AGI" means now. Right on schedule.
ARC Prize
OpenAI's GPT-6 Astra on ARC-AGI-3 | ARC Prize
OpenAI's GPT-6 Astra reaches state-of-the-art results on ARC-AGI-3
⚡️ Six fully open models, 0.9B to 375B. IFM means it.
MBZUAI's Institute of Foundation Models just dropped K2 Horizon, a fleet of six models with full weights, code, and training data. Not "open" in the Llama sense. Actually open.
375B at the top, 0.9B at the bottom for on-device. Every size tuned for a specific workload.
Honestly the "world's largest fully open model" claim is real this time.
MBZUAI's Institute of Foundation Models just dropped K2 Horizon, a fleet of six models with full weights, code, and training data. Not "open" in the Llama sense. Actually open.
375B at the top, 0.9B at the bottom for on-device. Every size tuned for a specific workload.
Honestly the "world's largest fully open model" claim is real this time.
Institute of Foundation Models
Introducing K2 Horizon: Frontier Performance, Radically Open
Explore K2 Horizon, IFM’s open-source fleet of six frontier AI models for reasoning, coding, agentic workflows, edge devices, and enterprise deployment.
❤1🔥1
⚡️ Google is suspending accounts for using 3rd-party tools with Gemini CLI
Using proxies or wrappers that tap Gemini's OAuth to access Antigravity resources violates ToS and people are getting banned. Not just from the CLI. From their whole Google account.
OpenAI and Anthropic explicitly allow 3rd-party harnesses. Google locks you out of Gmail for the same thing.
Developers are noticing.
Using proxies or wrappers that tap Gemini's OAuth to access Antigravity resources violates ToS and people are getting banned. Not just from the CLI. From their whole Google account.
OpenAI and Anthropic explicitly allow 3rd-party harnesses. Google locks you out of Gmail for the same thing.
Developers are noticing.
X (formerly Twitter)
Gergely Orosz (@GergelyOrosz) on X
❤3
⚡️ Coding agents reach for grep. Not LSP. Every time.
Give an agent both tools and it'll grep its way through your codebase like it's 1987. LSP just... sits there.
Grep is all over training data, which likely explains it. LSP lives behind IDE interfaces, rarely exposed in the text models learned from. So agents default to what they know.
It works. But it's a ceiling. Source
Give an agent both tools and it'll grep its way through your codebase like it's 1987. LSP just... sits there.
Grep is all over training data, which likely explains it. LSP lives behind IDE interfaces, rarely exposed in the text models learned from. So agents default to what they know.
It works. But it's a ceiling. Source
AgentConnect
Grep beats LSP? Why coding agents ignore your fancier tools
I compared grep with LSP-backed semantic navigation across code-finding and editing tasks. The results show why a tool's LLM-friendliness may matter as much as the capability behind it.