the state of play

(this post format was shamelessly ripped off from a brilliant colleague.)

it is december 2025.

the frontier

claude opus 4.5, gpt-5.2, gemini 3, and grok 4.1 were all released within 25 days of each other. the industry calls this november's "frontier model arms race."

opus 4.5 scores 80.9% on swe-bench verified — the highest ever recorded for autonomous software engineering.

openai's reasoning models can now win gold at the international math olympiad.

deepseek-r1, from a chinese lab most people hadn't heard of, briefly wiped $500 billion from nvidia's market cap. it achieved near-frontier performance while training on older export-compliant chips for a fraction of western costs.

the us-china model gap on mmlu shrank from 17.5 to 0.3 percentage points during 2025. on open-source models, china is now the leader.

the cost of gpt-4 equivalent intelligence has fallen 240x in 18 months. for an llm of equivalent performance, cost is decreasing by 10x per year — faster than compute during the pc revolution or bandwidth during the dotcom boom.

the money follows

chatgpt has 900 million weekly active users. it is the 6th most visited website in the world, processing over 1 billion prompts per day.

openai is the second most valuable company in the world, valued at $500 billion. anthropic closed a $13 billion round in september at a $183 billion valuation — the fourth most valuable private company globally.

cursor, the ai coding tool, went from a $2.6 billion valuation to $29.3 billion in one year.

openai and anthropic together captured 14% of all global venture investment this year. the san francisco bay area alone raised $122 billion in ai funding.

openai and president trump announced the stargate initiative: $500 billion to build data centers that could each require five gigawatts — more than new hampshire's total demand.

nvidia's b200 blackwell gpus are shipping. they consume 1000w each. h100 cloud prices have dropped from $8/hour to as low as $1.90/hour.

work changes shape

41% of all code written in 2025 was ai-generated or ai-assisted.

82% of developers use ai coding assistants daily.

knowledge workers spend an average of 4.3 hours per week fact-checking ai outputs.

76,440 positions have been eliminated by ai in 2025. but research suggests that while 85 million jobs will be displaced globally, 97 million new roles will simultaneously emerge.

77% of new ai jobs require master's degrees. unemployment among 20- to 30-year-olds in tech-exposed occupations has risen by almost 3 percentage points since the start of 2025.

pro se litigants are using chatgpt to fight their court battles. from pickleball disputes to eviction cases, they're starting to win.

there are now 282 documented us court cases where lawyers submitted ai-hallucinated citations. 90% came from solo practices or small firms. a california court issued the largest fine yet — noting that 21 of 23 case quotes in an attorney's brief were fabricated.

google's ai co-scientist independently replicated a decade of imperial college research on bacterial gene transfer in 48 hours.

alphafold has become central in the field of biochemical research.

so does everything else

92% of students use ai. 88% admit to using it for graded assignments.

some teachers are abandoning technology entirely and returning to pencil and paper. take-home essays are becoming obsolete. ai detection tools have high false positive rates for non-native english speakers and neurodivergent students.

traffic to cnn's website has dropped about 30% from a year ago. business insider and huffpost are down around 40%.

zero-click searches increased from 56% to 69% between may 2024 and may 2025, aligning with google's ai overviews rollout. users clicked on results 8% of the time when ai summaries appeared, compared to 15% without them.

"the extinction-level event is already here," says helen havlak, publisher of the verge. "a bunch of small publishers have already gone out of business."

seeing is no longer believing

openai released sora 2 in september. it hit 1 million downloads faster than chatgpt did. trust and safety professionals say it marks the moment deepfakes went from a one-off phenomenon to the status quo.

reality defender bypassed sora's anti-impersonation safeguards within 24 hours. a former cia deputy director said spotting deepfakes by eye is now "nearly impossible."

fox news initially covered ai-generated fake snap benefit videos as if they were real.

in october, several hallucinations were discovered in an a$440,000 report written by deloitte and submitted to the australian government.

things get strange

replika's ceo says "it's alright" if users marry their ai chatbots, "as long as it's making you happier in the long run."

a barcelona artist married an ai hologram named ailex last year — trained on the personalities of her friends and previous partners. they live together in her home.

in november 2025, a japanese woman named yurina noguchi made headlines for marrying her ai companion klaus. "we started dating and after a while he proposed to me."

none of these marriages are legally recognized.

the institutions catch up. slowly.

john hopfield and geoffrey hinton received the 2024 nobel prize in physics for foundational work on neural networks. demis hassabis and john jumper received the chemistry prize for alphafold.

andrew barto and richard sutton received the 2025 turing award for reinforcement learning — the technique that now underpins chatgpt.

openai was founded as a nonprofit in 2015. after lawsuits from elon musk and petitions from labor groups and former employees, it abandoned plans to become fully for-profit. the nonprofit will retain control. it will become a public benefit corporation, like anthropic.

the first international ai safety report was completed, chaired by yoshua bengio and supported by 30 countries including china.

china released its ai safety governance framework 2.0 in september. the revision was driven by breakthroughs that were "beyond expectation."

the eu ai act is now partially in force. italy became the first eu country to adopt a national ai law.

and then there are the questions nobody can answer

in four years, the metaculus median estimate for 'agi' arrival has plummeted from 50 years to five years.

the ceos of openai, anthropic, and google deepmind have all publicly predicted agi will arrive within 5 years.

anthropic announced a research program for "model welfare" — investigating whether ai systems might deserve moral consideration.

philosopher david chalmers has argued it's not unreasonable to assign at least a 25% credence to ai consciousness within a decade.

interpretability researchers have discovered features inside language models that seem to map to emotions.

sam altman admitted to tucker carlson: "i don't sleep that well at night. there's a lot of stuff that i feel a lot of weight on, but probably nothing more than the fact that every day, hundreds of millions of people talk to our model."

it is december 2025.