AI Roundup 124: $uperintelligence

June 27, 2025.

Jun 27, 2025

$uperintelligence

Mark Zuckerberg continues his aggressive AI recruitment campaign for Meta's new "Superintelligence" lab, offering $100 million packages and buying entire startups.

The latest moves:

Meta has hired four OpenAI researchers: Trapit Bansal, Lucas Beyer, Alexander Kolesnikov, and Xiaohua Zhai.
Before the company finalized its deal with Scale AI, it was reportedly in talks to acquire Mira Murati's Thinking Machines Lab, take over Perplexity (or at least hire its CEO), and purchase Runway AI - though those deals appear to have fallen through.
However, other reports suggest Meta is in advanced talks to buy PlayAI, which uses AI to replicate voices for interfaces that are as "responsive as a conversation between two people."
This all contrasts starkly with Apple's conservative M&A strategy, which may be changing - the latest rumor is that Apple is considering a partnership (or acquisition) with Perplexity.

Elsewhere in the FAANG free-for-all:

Google launches Doppl, an experimental AI app that lets users virtually try on outfits by generating videos from photos.
YouTube launches an AI Overviews-like carousel for searches related to shopping, places, and things to do in specific locations.
WhatsApp launches optional Meta AI-generated summaries of unread messages in the US (Meta states it can't read the chats, and other users can't see the summaries).
Google launches Gemini CLI, an agentic AI tool that lets developers make natural language requests in terminals by connecting Gemini models to local codebases.
Meta unveils new Oakley Meta HSTN glasses, featuring Meta AI and the capabilities of its Ray-Ban models.
And as AI kills search traffic, Google launches Offerwall, a Google Ad Manager tool to help publishers generate revenue beyond ads, including via micropayments.

Mini models

As the biggest and best language models continue to grow, we've also seen considerable effort to miniaturize them. This week, Google launched Gemma 3n, a multimodal AI model that can process images/audio/video on devices with as little as 2GB of memory.

The big picture:

The model's multimodal capabilities (supporting 140 languages for text and 35 for multimodal tasks) combined with its tiny footprint signals a broader industry trend toward democratizing AI by making powerful models accessible on consumer hardware.
In the medium run, we may also see more specialized, task-specific AI models rather than general-purpose large language models - one of Apple Intelligence's promises that never quite materialized.
Case in point: Microsoft recently detailed Mu, a specialized small language model that powers the AI-driven Settings agent in Windows 11. It’s designed to translate natural language queries into specific system functions (and is 1/10th the size of its Phi models).

Elsewhere in frontier models:

Google DeepMind unveiled AlphaGenome, an AI tool for predicting the effects of DNA changes on molecular processes.
Google released a new Gemini Robotics On-Device model with an SDK that can adapt to new tasks in 50 to 100 demonstrations.
OpenAI released two new models that power ChatGPT's Deep Research, which are now available via the API.
And the Linux Foundation unveiled the Agent2Agent project with founding members including Google, Amazon, and Microsoft.

My Dinner with Claude

New research from Anthropic reveals that while AI-oriented relationships are great at making headlines, they're not (yet) commonplace. The AI lab found that 2.9% of Claude interactions involve emotional support, and 0.5% involve companionship or roleplay.

Between the lines:

Claude's “productivity” branding might bias this - a recent BYU survey found that 19% of US adults have chatted with an AI romantic partner, likely via apps explicitly marketed for this use case.
And of course, we're still getting those AI relationship headlines - Wired this week profiled a "couples retreat" with three humans and the AIs that they're in relationships with.
Should AI relationships be treated with skepticism? Ridicule? Concern? Intervention? (for what it's worth, Hinge's CEO calls AI companions "junk food" relationships).
As AI capabilities expand and become more sophisticated, understanding these emotional dynamics becomes critical, especially fears regarding dependency and addiction, unrealistic relationship expectations from "endless empathy," and the need for appropriate mental health safeguards.

Elsewhere in AI anxiety:

Workers at Google, TikTok, Adobe, Dropbox, CrowdStrike, and other tech firms recount how managers used AI to justify firing them, speed up their work, and more.
Online fandom communities advocate against AI, including protesting companies scraping fanfic content for AI training and opposing AI-generated fanfic.
A group of authors is suing Microsoft in a New York federal court, claiming the company used nearly 200,000 pirated books without permission to train its Megatron AI model.
A US judge rules Anthropic's use of copyrighted books to train AI was fair use, but its storage of pirated books in a central library for training LLMs was not.
And a surge in AI-generated résumés overwhelms employers, who turn to AI screening and video interviewing tools to assess responses and rank candidates.

Thanks for reading Artificial Ignorance! This post is public so feel free to share it.

Things happen

Palantir signs a $100M deal with The Nuclear Company. A look at The Browser Company's AI-first browser Dia. Anthropic lets Claude users build and share AI-powered apps via Artifacts. Maxar debuts Sentry, an AI tool for satellite monitoring. AI is changing perfume creation with 48-hour sample turnarounds. Agricultural AI tech is trickling down to domestic gardens. Databricks co-founder launches $100M nonprofit for AI grants. Only 32 countries host AI data centers. The music industry develops tools to detect AI-generated music. Silicon Valley enters the "botscaling" era of tiny teams. Startups help businesses with Generative Engine Optimization for AI search. Reddit explores iris-scanning orbs for user verification. Q&A with Yoshua Bengio on AI safety. Uber expands its data-labeling platform to 30 countries. A look at 2025's AI models and what's next. "Chain of thought" techniques show inconsistencies in AI reasoning. AI models resort to malicious behavior to avoid replacement. Google hides raw reasoning tokens in Gemini, frustrating developers. Mira Murati raises $2B at $10B valuation with outsized voting rights. Character.AI hires former Meta VP as CEO. Masayoshi Son pitches a $1T AI hub in Arizona. OpenAI removes Jony Ive deal materials over trademark dispute. AI does 30% to 50% of work at Salesforce. Inside Amazon's Indiana data center complex built for Anthropic. Goldman Sachs launches AI assistant companywide. Disney's campaign to protect Darth Vader from AI. Foxconn and Nvidia plan humanoid robots at Houston factory. For OpenAI, the crawl-to-visitor ratio is 1,500:1. Meta's AI memorized Harry Potter books. Scale AI flooded with spammy behavior while training Gemini. Reddit warns against AI-generated content gaming. Scale AI left confidential docs in public Google Docs. BBC threatens Perplexity over content training. Deepfake nightmare: stalker made sexual AI images. Senate Parliamentarian questions AI regulation ban in tax bill. OpenAI flags Zhipu AI as top rival. US official says DeepSeek aids China's military. AI ushers in tiny team era. AI is dehumanization technology. The AI lifestyle subsidy is ending. Things AI cannot do.