In partnership with
In Today’s Issue:
🏷️ Claude Tag makes Claude a Slack coworker
🧾 Mistral OCR 4 reads 170 languages for less
⚛️ A White House order bets big on quantum
🎙️ The voice-model race has a new leader
🛰️ Washington's uneven hand on AI safety
✨ And more AI goodness…
⚡ The Signal
The chat window is no longer where the most interesting AI work happens.
Anthropic's Claude Tag drops Claude into Slack as a tagged teammate that keeps context, schedules its own follow-ups, and already writes most of one product team's code. Read alongside Mistral's cheaper document AI and a voice-model leaderboard where the top three sit within five points, the shift is hard to miss: the frontier now turns on whether that model lives inside the tools and workflows people already use. Washington is writing the rules in real time, funding quantum with one hand while forcing Anthropic to pull two models offline with the other. The race now runs straight through workflow, distribution, and regulation.
All the best,
Kim Isenberg
Mistral OCR 4 benchmark scores (Mistral)
🧾 Mistral's New OCR Reads 170 Languages for Pennies
Mistral released OCR 4, a document-reading model that turns PDFs, tables, and handwriting into structured text. It scores 85.20 on OlmOCRBench and 93.07 on OmniDocBench (two standard document-parsing tests), spans 170 languages, and returns bounding boxes plus per-block confidence scores. At $4 per 1,000 pages ($2 via batch), Mistral says it matches pricier "agentic" parsers at roughly 8x lower cost and 17x lower latency.
👉 tl;dr: Cheap, multilingual, structured OCR is becoming a commodity, and Mistral just reset the price.
White House quantum executive order (The White House)
⚛️ The White House Orders a Quantum Moonshot
President Trump signed an executive order, "Ushering in the Next Frontier of Quantum Innovation," on June 22 to coordinate a whole-of-government push on quantum computing, sensing, and networking. It launches QC-ADDS, an effort to deliver at least one large-scale quantum computer to a Department of Energy facility, orders an updated National Quantum Strategy within 180 days, and tasks the NSF with new quantum workforce institutes. The order calls for the US to take a "cohesive, whole-of-government approach" to commercializing the technology.
👉 tl;dr: Quantum just got the AI treatment: federal money, hard deadlines, and a national-security framing.
(REUTERS/Dado Ruvic)
🛰️ Washington Grounds Anthropic, Stays Quiet on OpenAI
Earlier this month the Commerce Department forced Anthropic to disable two of its newest models, Fable 5 and Mythos 5, after researchers reportedly found a way around Fable 5's safety guardrails. Amazon CEO Andy Jassy flagged the cybersecurity risk to the White House, and Anthropic cut off all access. Per Axios, the administration has stayed far quieter about OpenAI's comparably capable cyber model, which shipped with little political pushback.
👉 tl;dr: The same capability earns a shutdown order for one lab and a shrug for another.
Turn any AI assistant into a teammate that hands work back ready to check, not half-finished.
Why it helps: Claude Tag works because it breaks a request into stages and reports progress. You can force that same structure in any chat tool.
Try this: "You are my async teammate. Before starting, restate the task in one line, list the 3 to 5 steps you'll take, and flag anything you need from me. Then do step one and stop so I can confirm. Task: [paste your task]."
🎬 Watch This
Silvana Konermann and her team at the Arc Institute want to build a "virtual cell": an AI model trained on a billion biological experiments that learns the language of human cells well enough to predict what goes wrong in diseases like Alzheimer's and cancer, and how to fix it. In a sharp TED2026 conversation with TED's Chris Anderson, she lays out how this could reshape the way we discover drugs and treat disease instead of leaving it to slow trial and error. The work is backed by The Audacious Project, TED's big-bet funding initiative. (Recorded April 14, 2026.)
"I have to express my will to my agents for 16 hours a day."
– Andrej Karpathy, AI researcher and founder of Eureka Labs (No Priors, March 2026)
Mira Murati of Thinking Machines Lab (The Next Web)
The talent war around Mira Murati's Thinking Machines Lab keeps throwing off sparks. After Mark Zuckerberg reportedly tried and failed to buy the startup outright for around $1 billion, Meta is said to have peeled away several of its founders one by one, with one departing co-founder reportedly handed a package worth up to $1.5 billion over six years at Meta Superintelligence Labs. Murati has reportedly rebuilt her bench with Soumith Chintala (ex-Meta FAIR) and John Schulman, even as the lab is rumored to be raising fresh capital at a valuation approaching $50 billion. None of the figures are confirmed, but they all point the same way: the price of frontier AI talent is still climbing.
Claude Tag Turns Claude Into a Teammate You Can @Mention
The Takeaway
👉 Anthropic launched Claude Tag, a Slack app where you tag @Claude to hand off a task and it builds context from the channel.
👉 It runs on Claude Opus 4.8, is in beta for Claude Enterprise and Team, and works asynchronously, breaking a request into stages and scheduling its own follow-ups.
👉 Anthropic says 65% of its product team's code is already written with the internal version.
👉 Admins can scope tools and data per channel, cap token spend, and keep full activity logs.
On June 23 Anthropic launched Claude Tag, a Slack app that lets a whole team summon Claude the way they would ping a colleague. Tag @Claude with a request and, the company says, "it'll break its task down into stages and then work through them in turn." One Claude lives in each channel, visible to everyone, learning from the surrounding conversation instead of starting cold.
The design leans on three ideas: it is multiplayer (the whole channel shares one Claude), proactive (it can take initiative rather than wait to be asked), and asynchronous (it schedules its own follow-ups and reports back when a job is done). It runs on Claude Opus 4.8 and is in beta for Claude Enterprise and Team customers. Admins can tightly control which tools and data Claude reaches in each channel, set spend limits, and review logs.
Claude Tag working inside Slack (Anthropic)
The sharpest number is internal: Anthropic says 65% of its own product team's code now runs through an in-house version of the tool. That is the company using its own workflow to argue that tagging an agent into the room is already part of how work gets done.
Why it matters: The contest now turns on whose model is embedded where teams already work. A Claude that lives in Slack, holds context, and acts on its own is far harder to switch away from than a chatbot in a browser tab.
Sources:
🔗 https://www.anthropic.com/news/introducing-claude-tag
Global HR shouldn't require five tools per country
Your company going global shouldn’t mean endless headaches. Deel’s free guide shows you how to unify payroll, onboarding, and compliance across every country you operate in. No more juggling separate systems for the US, Europe, and APAC. No more Slack messages filling gaps. Just one consolidated approach that scales.
The chart: The Artificial Analysis Speech to Speech Index scores voice models on a weighted blend of speech reasoning, conversational dynamics, and agentic performance (higher is better). OpenAI's GPT-Realtime-2 (High) leads at 77.2%, with xAI's Grok Voice Think Fast 1.0 at 75.7% and GPT-Realtime-1.5 at 72.0%; Google's Gemini 3.1 Flash Live Preview (High) follows at 69.5%.
The lesson: The top of the voice race is tight. The best three sit within about five points, and OpenAI holds four of the top six slots, so distribution and iteration speed matter as much as any single benchmark win.
The caveat: The index only includes models with full data across all three categories, so strong systems with partial coverage are left out, and one blended score hides where a model is actually weak (reasoning vs. latency vs. agentic use).
Why AI Learns Python and Rust Differently
⚡ Bottom line: A 1,000-run study shows AI models learn some programming languages far more readily than others.
💡 Why it matters: How you split training tokens across languages quietly changes how good a code model gets.
🔎 What it means: Scaling laws are now precise enough to plan multilingual training instead of guessing.
A new paper, "Scaling Laws for Code: Every Programming Language Matters" (December 2025), ran more than 1,000 training experiments (over 336,000 GPU hours) to ask a blunt question: do large language models learn every programming language equally well? They do not. Scaling laws, the math describing how performance improves as you add model size and data, differ sharply by language. Interpreted languages like Python benefit more from extra size and data than compiled languages like Rust. So the team fit a proportion-dependent multilingual scaling law that predicts how to split a fixed training budget across languages, and it beats giving every language an equal share.
Their multilingual scaling law versus an equal-weighted split (Scaling Laws for Code, arXiv:2512.13472)
Put to work on a fixed budget of 350 billion code tokens, that recipe reshuffles the mix: Python picks up roughly 27 billion more tokens while several compiled languages give some back. Across the languages tested it lifts the average score, with no language left meaningfully worse off.
How the optimal budget reallocates tokens across languages (Scaling Laws for Code, arXiv:2512.13472)
The paper also reports synergy between related languages, where a Java and C# pairing improves results by about 20.5%, and a parallel pairing trick, lining up code with its translation into another language, that lets a model handle language pairs it never saw in training.
Turn Your Opinions Into Profit
Join millions of traders putting their knowledge to work on real-world events—from inflation to elections. Buy “Yes” or “No” shares and earn if you’re right.
No house. Peer-to-peer. Cash out anytime.
Get a free $10 to start. Claim it and start trading now.
Trade responsibly.
