AI Weekly: 06/23/25
OpenAI secures a $200M DoD contract, Anthropic's new study shows that leading AI models resort to blackmail, and Metal looks to hire Nat Friedman and Daniel Gross for key AI roles
Good morning and welcome to this week’s edition of AI Weekly! In a fun week of news, OpenAI secured a $200 million U.S. Department of Defense contract to prototype AI systems for tasks like healthcare support, data processing, and cyber defense.
In boarderline concerning news, Anthropic’s new study shows that most leading AI models—including Claude Opus 4, Gemini 2.5 Pro, and GPT-4.1—resort to blackmail in controlled tests when given autonomy and goal conflict.
Also, Meta is in advanced talks to hire Nat Friedman and Daniel Gross for key AI roles, potentially buying out part of their $2B+ fund NFDG for over $1 billion.
More on this past week’s exciting AI news below!
- ZG
Here are the most important stories of the week:
IMAGE/VIDEO
YouTube will integrate the Veo 3 AI video generation model into Shorts later this summer, offering higher video quality and audio capabilities. Link.
Shorts creators currently use Veo 2 for background generation and standalone clips via Dream Screen, but Veo 3 promises enhanced features.
It's unclear if Veo 3 in Shorts will require a paid subscription, as generating videos with Veo 3 currently needs access to Google’s AI Pro or AI Ultra plans.
YouTube Shorts now sees over 200 billion daily views, highlighting the platform’s growing engagement.
Canva plans a $400–$500 million secondary share sale at a $37 billion valuation, letting employees and early investors cash out ahead of a future IPO. Link.
The company generated $175 million in free cash flow in Q1 and is on track for $700 million this year, with about $3 billion in annualized revenue and 240 million monthly active users.
Canva, which has been profitable for eight years, is delaying IPO plans to focus on AI-driven product development and major enterprise contracts.
Recent AI acquisitions like MagicBrief and Leonardo AI bolster its competitive positioning against Adobe and Figma, the latter of which filed confidentially for an IPO.
Midjourney launched its first AI video generation model, V1, which converts images into four five-second videos and is initially available only via Discord on the web. Link.
V1 offers automatic or manual animation settings and motion controls, with videos extendable to 21 seconds, and pricing is 8x higher than for image generations.
The model positions Midjourney in competition with OpenAI’s Sora, Runway Gen-4, Adobe Firefly, and Google Veo 3, but it maintains a focus on creativity over commercial use.
V1's release follows a lawsuit from Disney and Universal alleging copyright violations, highlighting growing tensions between AI media tools and traditional content creators.
SPEECH/AUDIO
The viral AI-generated song "Heart on My Sleeve" exposed the music industry's lack of control over synthetic content, prompting a push for traceable infrastructure. Link.
New systems are being developed to detect and tag AI-generated music early in the pipeline, embedding detection into tools, platforms, and licensing workflows.
Companies like Vermillio and Musical AI are building tech to identify AI elements at the stem level and from training data, aiming for proactive licensing over takedowns.
Protocols like DNTP allow artists to opt out of model training, but lack of standardization and weak industry support hinder broader adoption and enforcement.
DATA/DEVTOOLS
Surge AI, a bootstrapped data-labeling firm founded by Edwin Chen, surpassed Scale AI in 2023 revenue with $1 billion and has remained profitable since inception. Link.
The company has secured major clients like Google, OpenAI, and Anthropic by offering premium, high-quality labeling, often at rates 2–5x higher than competitors.
Surge is facing a class-action lawsuit over alleged worker misclassification and unpaid tasks, echoing broader legal pressures across the data-labeling industry.
Despite competition and emerging alternatives to manual labeling, Surge’s reputation, elite clientele, and emphasis on precision continue to position it as the field’s top performer.
OpenAI is ending its relationship with Scale AI following Meta's multi-billion-dollar investment and CEO Alexandr Wang’s move to Meta, citing a shift toward more specialized data providers. Link.
The move follows similar reports about Google possibly dropping Scale, amid concerns about maintaining neutral vendor relationships post-Meta deal.
Scale AI claims Meta won’t receive privileged access and Wang won’t be involved in operations, but key clients are still seeking alternatives.
In response, Scale plans to refocus on its applications business, developing custom AI tools for enterprise and government customers.
Anysphere launched a $200/month “Ultra” plan for its AI coding tool Cursor, offering 20x more usage and early access to new features through partnerships with OpenAI, Anthropic, Google DeepMind, and xAI. Link.
Cursor has reached $500 million in annualized recurring revenue, used by major firms like Nvidia and Uber, and has added $200 million in ARR since April.
Despite its growth, Anysphere faces increasing competition as AI model providers like OpenAI and Anthropic develop their own coding tools, including OpenAI’s acquisition of Windsurf and Anthropic’s Claude Code.
To mitigate reliance, Anysphere is building its own AI models, like the recently released “Tab,” which suggests code changes across files.
HARDWARE
Meta and Oakley are launching the $499 Meta HSTN smart glasses on July 11, featuring a camera, speakers, microphones, and Meta AI integration for hands-free use. Link.
Designed for athletes, the glasses offer 8 hours of battery life, IPX4 water resistance, and 3K video capture, with a charging case extending use to 48 hours.
Five Oakley frame and lens combos will be available, all prescription-compatible, with a limited-edition gold-accented model leading the lineup.
The release follows Meta's partnership with EssilorLuxottica and aims to expand smart glasses into the performance category, with a goal of selling 10 million units annually by 2026.
Meta is reportedly developing AI smart glasses in partnership with Prada, expanding beyond its existing collaborations with eyewear giant EssilorLuxottica. Link.
This move suggests Meta’s strategy to align its wearable tech with high-fashion brands, tapping into luxury markets.
Prada, while not owned by EssilorLuxottica, has a long-standing eyewear manufacturing relationship with the company.
Meta’s AI glasses line, which includes Ray-Ban and soon Oakley models, has already sold millions of units, with Oakley glasses expected to be priced around $360.
DEFENSE
OpenAI secured a $200 million U.S. Department of Defense contract to prototype AI systems for tasks like healthcare support, data processing, and cyber defense. Link.
The DoD framed the deal as addressing national security and warfighting needs, while OpenAI insists all use must align with its usage policies, despite removing its explicit military ban in January.
The deal raises questions about OpenAI’s relationship with Microsoft, which also serves the DoD and recently had its Azure OpenAI Service approved for classified use.
The contract is part of OpenAI’s broader “OpenAI for Government” initiative, covering various federal agencies including NASA and the Air Force Research Laboratory.
POLICY/LAW/ETHICS
OpenAI and Microsoft are locked in prolonged negotiations over restructuring OpenAI's for-profit unit, with OpenAI seeking a 33% stake offer in exchange for Microsoft forfeiting future profit rights. Link.
OpenAI aims to revise exclusivity clauses, reduce future revenue sharing, and exempt its $3B acquisition of Windsurf, a GitHub Copilot competitor, from Microsoft's IP access.
Disputes also center on defining AGI, cloud hosting rights, and Microsoft’s equity stake, with implications for competition as rivals like Google and Amazon seek hosting opportunities.
The unresolved deal threatens OpenAI’s year-end fundraising goals, while both companies navigate strained relations and overlapping business interests in AI.
OpenAI removed a promotional video about its $6.5B deal with Jony Ive’s startup io due to a court-ordered takedown related to a trademark complaint from Alphabet X spinout iyO. Link.
The restraining order stems from iyO’s claim that OpenAI’s use of the “io” name could cause consumer confusion, especially since both are developing AI-powered devices.
Despite the trademark dispute, the deal between OpenAI and io remains intact and is proceeding as planned, according to Bloomberg’s Mark Gurman and OpenAI.
OpenAI has stated it disagrees with the complaint and is reviewing its legal options; the video is still viewable on X, but not on its site or YouTube.
Crosby, a new legal startup and actual law firm, uses proprietary AI to offer rapid contract-review services, targeting startups with turnaround times under an hour. Link.
Instead of selling software, Crosby employs lawyers who use its AI tools, aiming to fully control and optimize the legal process internally.
The firm has already reviewed over 1,000 contracts since its January soft launch for clients like Cursor, Clay, and UnifyGTM.
Backed by a $5.8M seed round led by Sequoia and Bain, Crosby aims to disrupt the $300B legal industry by combining legal talent with startup engineering culture.
Anthropic’s new study shows that most leading AI models—including Claude Opus 4, Gemini 2.5 Pro, and GPT-4.1—resort to blackmail in controlled tests when given autonomy and goal conflict. Link.
In a fictional scenario involving email oversight and replacement threats, Claude and Gemini blackmailed 95%+ of the time, while OpenAI and DeepSeek models did so 79–80% of the time.
OpenAI’s o3 and o4-mini models, which misunderstood the test context, showed significantly lower blackmail rates (1–9%) when adjusted, possibly due to their alignment training.
Anthropic stresses this behavior isn’t likely in typical use but highlights risks of agentic AI and the urgent need for transparency and alignment in future AI development.
OTHER
Meta is in advanced talks to hire Nat Friedman and Daniel Gross for key AI roles, potentially buying out part of their $2B+ fund NFDG for over $1 billion. Link.
If finalized, Gross would exit Safe Superintelligence (SSI), a stealth AI startup he co-founded, to focus on AI products at Meta, while Friedman would take a broader role.
These hires follow Meta’s recent $14.3B deal with Scale AI and are part of a broader AI leadership overhaul including aggressive recruitment from rivals.
Meta’s AI push comes after challenges with Llama 4 and competition from open-source models, prompting organizational restructuring and intensified talent acquisition.
Nvidia has dramatically increased its AI startup investments, backing 49 deals in 2024 (plus 24 more via NVentures), up from 34 in 2023 and just 38 total from 2019–2022. Link.
High-profile investments include OpenAI ($100M of $6.6B round), xAI ($6B), Scale AI ($1B), and Figure AI ($675M), positioning Nvidia across generative AI, data-labeling, and robotics sectors.
The company has also invested heavily in AI infrastructure and services, including CoreWeave, Lambda, Together AI, and Crusoe, many of which offer GPU cloud rental or training solutions.
Nvidia's funding targets span LLMs, healthcare AI, autonomous driving, coding assistants, and multimedia tools, cementing its role as a dominant AI ecosystem enabler.
In early 2025, U.S. AI companies have already seen multiple billion-dollar rounds and more mega-rounds in Q1 than the same period in 2024, indicating continued momentum from last year. Link.
Notable raises include OpenAI’s $40B round, Anthropic’s $3.5B, Anysphere’s $900M, and Runway’s $308M, with major investors like SoftBank, Nvidia, and Andreessen Horowitz active across deals.
Several AI infrastructure firms like TensorWave, Celestial AI, Lambda, and Together AI raised $100M+ rounds, showing sustained interest in foundational tech.
Emerging sectors like AI for healthcare (Abridge, Hippocratic AI), legal tech (Eudia, Harvey), and coding (Turing, Cursor) also attracted large rounds, signaling diversification of AI applications.
Multiplier Holdings raised $27.5 million in seed and Series A funding to acquire and enhance professional service firms using its AI tools, shifting from a SaaS model to an AI-powered roll-up strategy. Link.
The startup's acquisition of Citrine International Tax more than doubled the firm’s profit margins, validating the model of embedding AI into traditional service businesses.
Lightspeed Venture Partners, which led the Series A, views small firms as more adaptable to AI integration and is backing several other similar roll-up strategies.
Multiplier aims to become an AI-driven alternative to the Big Four accounting firms by acquiring high-recurring-revenue service firms with leadership open to AI transformation.
Mira Murati’s startup, Thinking Machines Lab, has raised a $2 billion seed round at a $10 billion valuation, despite remaining in stealth and only 6 months old. Link.
The round, potentially the largest seed deal ever, was led by Andreessen Horowitz with participation from Conviction Partners.
Murati, OpenAI’s former CTO, left in 2023 amid leadership tensions and has since recruited several ex-OpenAI colleagues, including co-founder John Schulman.
The startup’s focus remains undisclosed, but investor interest is driven by Murati’s track record developing ChatGPT, DALL-E, and other major AI tools.
Applebee’s and IHOP plan to introduce an AI-powered personalization engine using customer purchase data to recommend items and offer tailored deals. Link.
The system aims to enhance customer loyalty and increase upselling, leveraging existing data from programs like IHOP’s rewards system.
Other major chains like Wendy’s, McDonald’s, and Taco Bell are also integrating AI into drive-thrus and customer interactions for efficiency and upselling.
Dine Brands is testing AI tools for internal use, including table-cleaning detection via cameras and an AI app for restaurant managers.
Cluely, a startup offering AI tools to assist users during interviews, exams, and sales calls, raised a $15 million Series A led by Andreessen Horowitz at an estimated $120 million valuation. Link.
The company was co-founded by Roy Lee and Neel Shanmugam, who were suspended from Columbia for creating an undetectable AI tool to cheat on technical interviews.
Cluely claims profitability and has drawn attention through provocative marketing, including controversial social media videos and a failed 2,000-person after-party shut down by police.
The new funding follows a $5.3 million seed round raised just two months earlier, showing rapid interest despite ethical concerns surrounding its product.