In partnership with
Today: Mistral ships a document-AI model that runs on your own servers, four senior researchers leave Google DeepMind for rivals, and the AI memory boom lifts Micron to a record quarter. Here are five developments worth knowing from the past few days.
A WORD FROM OUR SPONSOR
The best prompt engineers aren't typing. They're talking.
Power users figured this out early: speaking a prompt gives you 10x more context in half the time. You include the edge cases, the examples, the tone you want — because talking is fast enough that you don't skip them.
Wispr Flow captures everything you say and turns it into clean, structured text for any AI tool. Speak messy. Get polished input. Paste into ChatGPT, Claude, Cursor, or wherever you work.
89% of messages sent with zero edits. 4x faster than typing. Works system-wide on Mac, Windows, and iPhone.
Models & Products
Mistral releases OCR 4, a self-hosted document-AI model. Mistral introduced OCR 4, a document-understanding model that extracts structured text, tables, and paragraph-level bounding boxes across roughly 170 languages and ships as a single container so regulated organizations can run it inside their own infrastructure. The company reports a top score of 85.20 on the OlmOCRBench benchmark and says independent annotators preferred its output over other leading OCR and document-AI systems in a majority of head-to-head comparisons. OCR 4 is available through Mistral's API, Amazon SageMaker, and Microsoft Foundry. (Mistral) (MarkTechPost)
Industry & Business
Four senior researchers leave Google DeepMind for OpenAI and Anthropic. Over roughly a week in mid-to-late June, several prominent Google DeepMind researchers departed for competitors, including Gemini co-lead Noam Shazeer, who said he is joining OpenAI, and AlphaFold lead and 2024 Nobel laureate John Jumper, who is moving to Anthropic, along with coding-research lead Jonas Adler and pretraining researcher Alexander Pritzel. Reporting linked the moves in part to rivals' planned public offerings and to Google's effort to close a gap in agentic coding. Alphabet's market value fell by roughly $269 billion across trading sessions during the stretch. (TechCrunch) (TechTimes)
Micron posts a record quarter on AI memory demand. Micron reported fiscal third-quarter 2026 revenue of $41.46 billion, up from $9.30 billion a year earlier, and GAAP net income of $28.24 billion, driven by demand for the high-bandwidth memory used in AI accelerators. The company guided to roughly $50 billion in revenue for the current quarter and about 86% gross margin. Shares rose after the June 24 report. (Micron) (CNBC)
SK Hynix files for a U.S. listing tied to AI memory growth. SK Hynix filed a Form F-1 with the SEC on June 24 to list American depositary shares on the Nasdaq under the ticker SKHY, in an offering reported at up to about $29.4 billion, following a confidential filing earlier in the year. Bank of America, Citi, Goldman Sachs, and JPMorgan are named as lead underwriters, and trading is targeted to begin around July 10. The move would rank among the largest U.S. listings and reflects surging demand for the company's high-bandwidth memory. (SEC) (CNBC)
Research & Security
Researchers detail "Agentjacking," an attack that hijacks AI coding agents. Security firm Tenet described an attack class it calls Agentjacking, in which a crafted error report sent to a monitoring service such as Sentry is later surfaced to an AI coding agent as trusted system data, causing the agent to run an attacker's code with the developer's own privileges. In controlled testing the researchers reported an 85% success rate against leading coding agents and identified thousands of exposed configurations; Sentry acknowledged the issue but described it as difficult to fully resolve at the platform level. The findings, disclosed to Sentry in early June, drew wider coverage over the past week. (The Hacker News) (SC Media)
That's today's edition of AI Current. Reply with what you're seeing in AI, and forward this to a colleague who would find it useful. See you tomorrow.
A WORD FROM OUR SPONSOR
Your next 100 customers are already in Apollo
Find, reach, and close your perfect deals — without juggling five tools or hiring more reps.
Apollo gives you everything you need to build real pipeline, fast. From inbound to outbound, first touch to close.
All in Apollo.
A WORD FROM OUR SPONSOR
The #1 Time-Series Database Built on Postgres.
Columnar storage. No split architecture. No new query language. TimescaleDB extends Postgres with hypertables, Hypercore compression, and continuous aggregates so analytics run on live data.
CERN uses it to handle sensor data from the Large Hadron Collider.
One database. No pipeline.