TechLook
A rumor with no confirmed trigger just erased trillions from global markets — because the AI trade is wound so tight that everyone leaning the same way sold on the same whisper. Meanwhile, Anthropic brought agentic AI into the place most business gets done, L'Oréal skipped the SEO playbook and handed its data directly to OpenAI, and Hollywood proved it takes more than a subpoena to silence an industry.
Trillions Gone, No Cause Found
South Korea's KOSPI plunged 10% in a single day, triggering circuit breakers, as a wave of algo-driven panic selling spread from Seoul to New York with no confirmed trigger.
Here's everything you need to know:
- KOSPI fell 10% in a single session; Samsung dropped 12%, SK Hynix 13%.
- The Philadelphia Semiconductor Index fell ~8%, Micron dropped 13%, Nasdaq shed 3%.
- No confirmed cause emerged — theories range from HBM demand slowing, Nvidia cutting Rubin orders, retail margin calls, pension rebalancing, to a rumored Korean tax on paper gains.
- Analysts describe it as mechanical: algo selling, forced liquidation among leveraged retail traders, and institutional rebalancing — not a fundamental break.
- The selloff concentrated in the same names that powered the AI rally over the past 18 months.
The entire AI rally is one nervous, over-leveraged bet flinching at its own shadow. Every founder should read this as a risk-management signal: if your company's valuation, fundraising, or customer base depends on the AI trade continuing to climb, you're exposed to a correction that has nothing to do with your product. Build revenue models that survive a six-month capital winter.
Tag @Claude — Slack Becomes an Agentic Coworker
Anthropic launched Claude Tag, bringing the agentic capabilities of Claude Code and Cowork into Slack channels — where most business context, decisions, and tools already live.
Here's everything you need to know:
- Teams tag @Claude with a task, and the AI breaks it into stages, works through them using approved tools, and responds when done.
- Claude learns over time, builds context across channels, and features an ambient mode that follows up on tasks that have gone quiet.
- Andrej Karpathy called it the "3rd major redesign of LLM UI UX."
- Rolling out today in beta for Enterprise and Team customers.
- Administrators control token-spend limits and data sharing policies.
- An Anthropic employee posted four best practices for using Claude Tag in Slack.
Going from chat and desktop to Slack — where the workflows, documents, and conversations that define an organization already live — is the most natural next step for agentic AI. This release will likely hurt a lot of "agentic coworker" startups that were betting on Slack being the last integration they shipped, not the first one Anthropic built.
L'Oréal Hands OpenAI the Keys to Its Product Database
L'Oréal struck a direct partnership with OpenAI, feeding up-to-date product information into the models powering ChatGPT so users get answers from L'Oréal's own documentation rather than Reddit threads or wiki summaries.
Here's everything you need to know:
- When users ask ChatGPT about L'Oréal brands, the assistant pulls from L'Oréal's own product notes, not reviews, forums, or third-party sources.
- OpenAI insists the deal won't tilt results toward L'Oréal over competitors.
- The partnership also brings OpenAI's image model into L'Oréal's CreAItech system.
- Maybelline virtual try-on will be available inside ChatGPT.
- L'Oréal effectively bypassed the entire GEO (generative engine optimization) industry that emerged to help brands game answer engines.
An entire industry spent the last year building playbooks to manipulate AI-generated answers. L'Oréal ignored all of it and just handed OpenAI the database directly. The real strategy isn't tricks — it's having the deal. For any brand selling AI-answerable products, the question is no longer "how do we rank well in AI search?" but "whose database is already inside the model?"
Hollywood Dropped a Film About Sam Altman — No One Had to Ask
Amazon MGM abruptly dropped "Artificial," Luca Guadagnino's nearly finished drama about Sam Altman's 2023 firing, killing its Oscar-qualifying run and SXSW slot — shortly after Amazon invested $50 billion in OpenAI.
Here's everything you need to know:
- Amazon MGM dropped the film; Netflix, A24, Focus Features, and Warner's Clockwork all passed too.
- Only Neon and Mubi reportedly remain interested.
- Amazon had just committed $50B to OpenAI before making the decision to abandon the film.
- A24 separately took $75M from Google's DeepMind for AI-filmmaking research.
- No edict was issued, no blacklist published — studios simply read the room and self-censored.
No edict, no blacklist — studios simply read the room and self-censored. The scariest kind of power doesn't have to censor anything; it just writes enough checks that everyone censors themselves. For founders building in any space where a single giant customer or investor has conflicting interests, the lesson is the same: independence isn't about what you're allowed to do — it's about what you can afford to lose.
Meta's $299 Glasses — No Ray-Bans Required
Meta launched a new "Meta Glasses" line starting at $299, dropping the Ray-Ban branding for a more affordable two-tier strategy spanning 26 styles and three designs.
Here's everything you need to know:
- Three designs: Meta Adventurer, Meta Fury, and Meta Glasses by Kylie (at $399 with Kylie Jenner's voice option for Meta AI).
- Built with EssilorLuxottica, powered by Muse Spark AI.
- Promises smarter answers, visual understanding, turn-by-turn navigation, and live translation.
- Compatible with prescription lenses, 8+ hours battery life.
- Drops the Ray-Ban and Oakley branding to hit a lower price point.
- Meta now holds ~80% of the AI glasses market.
The two-tier strategy — Ray-Ban for fashion credibility, Meta Glasses for price accessibility — is designed to extend Meta's lead before Google's rival hardware gets to market. For founders building on AR/wearable platforms, the question is whether Meta's dominant share makes it the de facto platform or just the first one consumers will leave when real competition arrives.
DuckDuckGo Up 75% — Google's AI Mode Is Pushing Users Away
DuckDuckGo installs jumped as much as 75% since Google I/O, as a large AI Mode box now dominates Google's mobile search app and ~68% of Google searches end without clicking through to an outside site.
Here's everything you need to know:
- DuckDuckGo installs rose up to 75% since Google's I/O update pushed AI Mode front and center.
- Google still controls ~90% of search, but user behavior is measurably shifting.
- ~68% of Google searches now end without a click to an external site (zero-click searches).
- The AI Mode box inside Google's mobile app is nearly the same size as the regular search box.
- Some users want chatbot answers, some want links, some want a quiet search page — Google is forcing the first option on everyone.
Google isn't losing search overnight, but it's making users feel pushed. For any business that depends on search traffic for customer acquisition, this is a slow-moving crisis. Media companies are already planning for a world where Google search traffic is unreliable. Founders should model their acquisition channels as if zero-click search is the default in 18 months.
Superhuman Buys GPTZero — Accelerator Meets Guardrail
Superhuman acquired GPTZero, pulling AI-detection and guardrail technology into the inbox productivity platform — creating an awkward tension between the accelerator and the brake.
Here's everything you need to know:
- Superhuman sells AI writing assistance (including the recently criticized AI feedback feature that mimicked real writers' voices).
- GPTZero sells AI detection and guardrails for educational and professional use.
- The plan is to fold GPTZero into Superhuman Go while keeping the education vertical as a core audience.
- The strategic question: will users trust a guardrail when the same company sells the thing the guardrail is supposed to guard against?
The tension is obvious and isn't going away. For founders building in the AI tooling space, this deal is a signal that the "detect vs. generate" separation is collapsing — and if you're not thinking about how those two functions relate inside your own product, someone else will decide for you.
One Person, Five Months, a 3,500-Year-Old Script Decoded
Tom Di Mino, a self-taught AI engineer and amateur linguist, systematically deciphered Linear A — the Minoan script that resisted academic linguists for 70 years — working alone with Python scripts written via an AI agent.
Here's everything you need to know:
- Di Mino produced a 408-entry lexicon, 40 phonetic values, and a nine-page grammar.
- He worked alone for five months from New York's Hudson Valley.
- He cross-referenced digitized corpora using Python scripts generated with an AI agent.
- Linguists at Cambridge and Rutgers are still reviewing his proposed phonetic values.
- Cracking an unknown script previously required a funded university department, decades of person-hours, and a team.
Di Mino had seven years of self-study, a laptop, and an AI agent — matching the throughput of an entire academic research program. One person now wields corpus-processing muscle that used to define a department. This is what "software is eating the world" looks like when the software writes itself.
Stanford's Proto Makes 120 Biology Models Play Together
Stanford's Brian Hie released Proto, an open framework that lets developers compose AI biology models across DNA, RNA, proteins, and ligands — something previously impossible due to incompatible software and mismatched dependencies.
Here's everything you need to know:
- Over 120 AI biology models exist, but combining them was impractical due to conflicting dependencies and incompatible APIs.
- Proto provides a shared interface for composing models across DNA, RNA, proteins, and small molecules.
- In tests, it designed cell-line-specific splicing patterns with 32% success testing 65 candidates vs. 7% using previous methods testing ~1,000.
- AI agents (including Claude) can write Proto programs directly.
- Nvidia simultaneously launched the BioNeMo Agent Toolkit, giving agents callable tools for protein structure prediction and molecular docking.
Proto is the integration layer that lets biological AI tools finally work together. If it becomes the standard interface, every new biology model plugs straight in — and the bottleneck shifts from "which model is best" to "what biological question do we want to answer."
⚡ Quick Hits
- OpenAI Bidi 1: Leaked clips show the upcoming voice model freestyling raps and singing — a leap beyond current text-to-speech in spontaneous tone and rhythm.
- Cloudflare CEO: Matthew Prince warned AI could "destroy small businesses" by making it harder for them to persuade AI shopping agents to buy their products.
- OpenArt Director: "Vibe directing" arrives — describe, generate, and edit consistent-character AI videos up to 5 minutes in a single conversation. Built by two ex-Googlers.
- Wikipedia: Jimmy Wales says AI can't edit articles directly (hallucinations remain too severe), but agents can flag overlooked stories for human editors.
- CIOs on the line: Fortune reports 74% of CIOs fear job risk if measurable AI ROI fails to materialize — careers are now directly tied to proving AI returns.
- Open-source beats Fable 5 on web design: A viral post (1.5M+ views) showed an open-weight model outperforming Anthropic's Claude Fable 5 on web design — pressure on proprietary labs keeps building.
- Gemini fixes Sheets errors: Google's AI now directly fixes spreadsheet formula errors inline, pulling context from your entire sheet.
- Trump admin presses Meta: Administration pushing Meta to submit models for government reviews — security scrutiny of advanced AI is escalating.
- AoFrio's 3.2M coolers: IoT-powered agentic AI — beverage coolers streaming temperature, location, and shelf data to predict failures and trigger repairs.
- Krea open-sources Krea 2 Raw and Turbo: Released an undistilled image model for fine-tuning alongside a fast 2K image generator for consumer hardware.
- AWS Summit DC: June 30 — 350+ sessions on agentic AI, security, and modernization. Free to attend.
Techlook — AI & tech signal for founders and builders.