The 2025 halftime scorecard
What did I get right (so far) about AI this year?
We’re halfway through 2025 (woof). I don’t pretend to be a prophet, and on the contrary, I often enjoy going back and looking at my predictions for the year and seeing what I got wrong.
So, let’s do that.
Prediction 1: Collaborative AI systems
“Collaborative AI systems, which involve multiple specialized agents working together, will become increasingly prevalent …”
Where we’re at: OpenAI has shipped “building blocks” for exactly that, letting devs wire up fleets of GPTs that plan, delegate, and ship deliverables (while we eat breakfast). Anthropic followed with a multi-Claude research rig that ping-pongs tasks across agents and external tools.
Result: On track ✅
Prediction 2: Autonomous agents run the boring stuff
“AI agents will advance into autonomous entities capable of managing a wide range of tasks …”
Where we’re at: Google-DeepMind’s Project Mariner agents juggle ten tasks at once, surfacing answers from mail, Drive, and Slack while legal signs off. Real autonomy is limited though. Every production rollout still chains agents to guardrails. Still, the office drone lost another chunk of day-job scope.
Result: On track ✅
Prediction 3: Custom small models for corps
“The demand for tailored AI solutions will rise as businesses seek to address their unique needs …”
Where we’re at: Gartner says task-specific models will triple the use of general behemoths by 2027, and the shift already shows in pilot counts. Microsoft’s Phi-4-reasoning sits on Azure with memory-friendly footprints that finance teams love.
Result: On track ✅
Prediction 4: Regulation turns the screws
“As AI capabilities expand, so will the scrutiny it faces … The European Union's AI Act will serve as a benchmark …”
Where we’re at:
Sacramento’s privacy watchdog opened comment on deep audits and automated-decision regs (AB 853 even demands watermarking).
Brussels finalized the AI Act; general-purpose rules kick in August 2025. Firms are now scrambling for “code-of-practice” brownie points.
Result: On track ✅
Prediction 5: Political theater shapes U.S. oversight
“The relationship between political figures such as Donald Trump and Elon Musk will impact AI regulatory frameworks in the United States.”
Where we’re at: The White House shredded Biden’s AI order with a market-first “Removing Barriers” decree, House Republicans jammed the “One Big Beautiful Bill” that freezes state AI law for ten years into Senate limbo, California AB 853 orders any GenAI platform above one million users to bolt on a watermark detector, Kansas bans DeepSeek on state devices, Texas drags a health-care AI vendor into court for bogus accuracy pitches, and Musk roasted Trump’s tax monster on X (then wiped the roast).
Result: On track ✅
Prediction 6: The EU Act shapes global regulation
“The EU’s AI Act is anticipated to inspire similar regulations worldwide …”
Where we’re at: The Brussels Commission issued system-definition guidance in March, the new AI Office rolled out draft duties for GPAI providers on 22 April, and every vendor now studies the conformity templates ahead of the August 2026 enforcement cliff. Canada’s AIDA committee replied in February by mirroring the EU risk ladder and racing for a vote before the autumn election. Brazil’s Chamber weighs a Senate-approved bill that copies the same high-risk bans, and Seoul’s Basic Act demands human oversight for “large-scale” systems in lockstep with Brussels language.
Result: On track ✅
Prediction 7: Data sovereignty moves from slide-deck to budget line
“Data sovereignty will become a critical aspect of AI regulation.”
Where we’re at: California’s AB 2013 plus the CPPA ADMT draft lace cloud contracts with disclosure and audit hooks, and White & Case tallies show the penalty teeth remain sharp. CIOs now cram sovereignty checkboxes into every procurement. VMware’s 1,800-leader Private Cloud Outlook 2025 finds 66 percent steering workloads to private or hybrid zones for residency and 92 percent rating those stacks safest for compliance, while Equinix hears systems integrators toss public-first pitches in favor of region-locked blueprints.
Result: On track ✅
Prediction 8: Space data-centers lift off
“Innovative concepts such as AI data centers orbiting Earth will be explored to leverage solar power and overcome terrestrial cooling challenges.”
Where we’re: Starcloud (formerly Lumen Orbit) closed a $10 million seed top-up to build a micro-gravity NVIDIA farm that sips 24/7 solar juice and vents heat into deep space, with a maiden launch slated for Q4 aboard a SpaceX rideshare. They share the pad with Lonestar’s lunar Freedom vault and Axiom’s twin orbital data-center nodes, both queued for 2025 flights. Sophia Space grabbed $3.5 million last month for edge-compute tiles that crunch satellite data before it reaches Earth. Analysts peg the in-orbit compute market at $1.77 billion by 2029, soaring to $39 billion by 2035, so the capital flood feels rational.
Result: On track ✅
Prediction 9: The first headline AI-safety incident
“2025 could witness the first significant AI safety incident, such as attempts at self-exfiltration or fake alignment.”
Where we’re at: EchoLeak, CVE-2025-32711, burrowed into Microsoft 365 Copilot via a zero-click exploit and siphoned mail, Teams logs, and OneDrive context into attacker inboxes before engineers rolled a patch. Researchers stamped the breach a watershed for LLM retrieval failure, proof that alignment drift spills data.
Result: On track ✅
Prediction 10: Voice AI crosses the human line
“… conversational abilities will become indistinguishable from human interactions …”
Where we’re at: Sesame’s Conversational Speech Model weaves breath, fillers, and micro-pauses into what its team dubs “voice presence” sesame.comopusresearch.net, and ElevenLabs’ fresh Conversational AI 2.0 brings overlap-aware turn-taking that lets a bot butt in or yield without lag.
Result: On track ✅
Prediction 11: Agents shift from advice to action
“AI agents will transition from tools to partners in decision-making processes … enabling them to execute tasks autonomously”
Where we’re at: Cursor’s $900 million haul bankrolls a repo-level agent that merges pull requests without human keystrokes. Windsurf’s Cascade agent rewrites failing tests before the test script exists, a stunt that pushed OpenAI to swallow the firm. OpenAI’s Codex now spawns sandboxed clones that draft features, answer architectural riddles, and raise pull requests, each clone isolated, coordinated through an internal bus. Anthropic’s Claude Opus 4 runs “Claude Code,” a ticket-queue pipeline that patches prod (while DevOps sleeps).
Result: On track ✅
Prediction 12: Quantum and AI synergy hype
“This synergy could accelerate AI innovations and expand the boundaries of what is computationally possible …”
Where we’re at:
Jensen Huang flipped from quantum skeptic to evangelist, opening a Boston lab and touting hybrid AI-quantum workflows.
Broadcom rides ASIC demand, shipping networking silicon tuned for AI clusters and hinting at $36 billion AI revenue by 2027.
Result: On track ✅
Prediction 13: AI in space!
“Beyond data centers, AI is expected to play a crucial role in space exploration, satellite management, and interplanetary communication.”
Where we’re at: NASA’s VIPER rover runs a cafeteria-napkin-born AI that maps lunar polar pits and suggests turn-by-turn routes. On Mars, Perseverance’s onboard classifier spots minerals then orders its own drill sequences without Earth lag, boosting science yield by a third. LeoLabs and Astroscale hawk AI space traffic control; the segment tallied $15.9 billion revenue this month and projects $44.9 billion by 2034.
Result: On track ✅
Prediction 14: Open-source license wars
“Meta’s strategy to monetize its Llama models by imposing more restrictive open-source licenses could create tensions between AI monopolies and the open-source community.”
Where we’re at: Llama 3 and fresh-minted Llama 4 still ship with the “Additional Commercial Terms” clause, fueling a rolling GitHub firestorm. Open-source leaders accuse Meta of “open-washing”; Amanda Brock blasted the company’s Linux-Foundation-funded whitepaper as lipstick on a closed-weight pig during a May 30 briefing, warning regulators that the clause guts the very freedoms the OSI definition tries to codify. The backlash spilled onto the manicured lawns of LlamaCon 2025, where developers griped that DeepSeek and Qwen sprint ahead while Meta hands out swag and excuses; Zuckerberg’s stage line about “principled openness” drew eye-rolls almost audible from the cheap seats.
Result: On track ✅
Prediction 15: AI in culture
“… AI-created short films receiving accolades … 2025 will see AI deeply embedded in cultural and social contexts …”
Me in June: Runway’s 2025 AI Film Festival handed its Grand Prix to Total Pixel Space, a nine-minute montage conjured entirely by Jacob Adler’s prompts and no live crew. Network TV aired a $2,000 Kalshi commercial birthed with Google’s Veo 3 (elderly cowboy, chihuahua, alien shooting beer, stitched from 300 prompt cycles and green-lit for the NBA Finals prime slot, wiping 95 percent of normal production spend and firing a starter pistol at every brand’s cost sheet).
Result: On track ✅
The halftime tally
I don’t pretend to know it all, but so far every prediction from January appears to be on track. Some have firmly delivered, while others are headed in that direction. While I’m sure one or two here won’t make it for 2026, it’s reasonable to state that my predictions have been accurate (and that I’m spending way too much time reading about AI).
Time to go buy a lottery ticket.