Good morning. It's Monday, May 4, and we're covering DeepSeek’s rising AI costs and intensifying competition, the Pentagon’s push to expand AI partnerships across vendors, and a Chinese court ruling against AI firings.
Know someone who'd love this? Forward it their way. First time reading? Sign up here to keep up with the future of tech.
Pentagon Expands AI Deals
After a court fight with Anthropic over restrictions, the U.S. Department of Defense signs new AI agreements to avoid dependence on a single provider. It inks deals with NVIDIA, Microsoft, Amazon Web Services, and Reflection AI to deploy models and hardware on classified networks for operational use. Systems will run in IL6 and IL7 environments, the highest security tiers, as over 1.3 million personnel already use GenAI.mil for non-classified tasks.
Salesforce Introduces Agentforce Operations
Enterprise AI failures increasingly stem from flawed workflows, not weak models, exposing a structural gap as companies deploy agents at scale. Salesforce launches Agentforce Operations, a platform that restructures business processes into deterministic task sequences for specialized agents rather than relying on probabilistic decision-making. The system lets users upload workflows or use predefined blueprints, breaking them into explicit steps while adding observability through session tracing and optional human checkpoints.
Chinese Court Limits AI Firings
A Chinese court rules that replacing workers with AI does not qualify as a valid reason for dismissal, challenging a growing global automation trend. The Hangzhou Intermediate People's Court sides with an employee fired after his role was automated, finding the company failed to prove termination was necessary under labor law. The worker, earning 25,000 yuan monthly, was offered a demotion with a 40% pay cut before being dismissed, with arbitration awarding higher compensation than the initial 311,695 yuan ($45k USD) severance.
Benchmark Finds AI Creativity Splits
A new evaluation shows no single AI model dominates creative work, with performance varying sharply by task rather than overall capability. The Human Creativity Benchmark tests leading models across ideation, mockups, and refinement using expert human scoring. Results show Anthropic’s Claude leads ideation, Google DeepMind’s Gemini excels in design systems, and OpenAI’s GPT performs strongest in refinement.
POWERED BY ADOBE
The New Creative Agent in Adobe Firefly
Adobe just launched the public beta of Firefly AI Assistant in Adobe Firefly. Describe what you want to create and the assistant handles the rest, orchestrating multi-step workflows across Photoshop, Premiere, Firefly, Express, and more. You call the shots — defining the vision, shaping the details, and choosing what works — while the assistant accelerates execution.
VIDEO
Reacting to “Why AI Is So Smart but Also So Dumb?”
Karpathy explains AI’s “jagged intelligence,” rapid leaps in coding, and the shift to agent-driven software—why models excel at verifiable tasks but fail elsewhere.
COMPETITION
DeepSeek V4 Underscores Rising Costs and Intensifying AI Competition
DeepSeek released its v4 model on April 24, 2026, delivering performance close to leading U.S. systems from three to six months earlier and offering significantly lower pricing for customers. But unlike its earlier breakthrough, the model appears far more expensive and time-consuming to build, with no disclosed training cost and a 16-month development gap. Competition inside China has intensified, with rivals like Alibaba (Qwen), ByteDance (Doubao), Moonshot AI, and Z.ai crowding the field, while focus shifts on monetizable AI applications rather than raw model performance.
Government intervention—including pressure to use domestic chips from Huawei and blocking foreign deals—has further complicated development. The muted response suggests that efficiency gains alone no longer guarantee impact in a market where ecosystems, applications, and policy constraints increasingly shape outcomes. → Read the full article here.
FORWARD FUTURE ORIGINAL
Scott Santens Launches AI Pledge for Humanity to Call Tech Leaders' UBI Bluff
For a decade, AI leaders from Elon Musk to Demis Hassabis to Mustafa Suleyman have warned that universal basic income will be necessary as AI displaces workers — and almost none have actually funded it. Longtime UBI advocate Scott Santens just launched the AI Pledge for Humanity to close that gap: a two-tier public commitment where resource-holders pledge a percentage of income or wealth to UBI pilots and advocacy, while everyone else can sign a petition in solidarity.
Santens credits Sam Altman ($14M into a three-year pilot) and Jack Dorsey (tens of millions through #startsmall) as the rare exceptions who've put real money behind the rhetoric. Founding signatories include Andrew Yang and Dustin Moskovitz, with the list growing across multiple countries. → Read the full article here.
INDEPENDENCE
Replit CEO Signals Independence As AI Coding Market Heats Up
Amjad Masad said Replit aims to remain independent despite industry consolidation, including reports that rival Cursor may be acquired by SpaceX for $60 billion. Masad highlighted Replit’s rapid growth—from $2.8 million in 2024 revenue to a projected billion-dollar run rate—and strong unit economics, contrasting with competitors reportedly operating at negative margins.
He also pointed to high customer expansion, with net revenue retention reaching up to 300%, driven by demand for end-to-end AI app-building tools. At the same time, Masad accused Apple of unfairly blocking Replit’s App Store updates, suggesting the dispute could escalate to court. The remarks underscore intensifying competition among AI developer platforms, where profitability, platform control, and distribution are becoming as critical as model performance. → Read the full article here.
NEWS
What Else is Happening
AI Flags Pancreatic Cancer Early: Mayo Clinic model detects subtle CT scan changes up to three years before diagnosis, outperforming radiologists threefold; clinical trials test potential to enable earlier interventions.
Meta Buys Robotics AI Startup (Paywall): Meta acquires Assured Robot Intelligence to advance humanoid robots that understand and adapt to human behavior, signaling deeper investment in embodied AI beyond software.
Codex Gains Passive Activity Memory: OpenAI’s experimental “Chronicle” feature lets Codex capture users’ screen context to generate persistent cross-application memory, pointing toward more proactive AI assistants.
Anthropic Explores UK AI Chips Deal (Paywall): Anthropic holds early talks with London startup Fractile to secure efficient inference chips, aiming to ease compute shortages, lower costs, and diversify beyond NVIDIA suppliers.
NVIDIA B300 Servers Hit $1M in China: Prices for NVIDIA’s top AI servers nearly double U.S. levels amid export curbs and smuggling crackdowns, as Chinese firms scramble for scarce compute to power surging AI workloads.
Tweets
This is the first time I've ever seen an LLM operate a GUI as fast as a person, and it's surreal. pic.twitter.com/5kjwGMDpvd
@arixView post on X
