Diogo Almeida was one of the fathers of ChatGPT, but he believes his creation has failed to deliver.
asserted
creation → believe → ChatGPT
“It’s not leading to an AI-based economic revolution, which is what the early people really wanted,” he told an audience earlier this year.
asserted
he → lead → audience
Rather than simply lament his failure, Almeida has done something about it.
asserted
Almeida → lament → it
In the last week, Jev – his low-cost AI system – has become a viral sensation, prompting rumours of a $US10 billion ($14.4 billion) valuation, as well as a string of copycats.
asserted
Jev → become → copycats
Jev’s developer, TypeSafe, says the system is both hundreds of times cheaper and faster than the “frontier” models developed by OpenAI and Anthropic, and designed for use by machines, rather than humans.
asserted
system → say → humans
Almeida – recognisable by a trademark pink fluffy coat that he says he purchased on Temu for $US20 – believes Jev could displace ChatGPT and Claude, and upend the two trillion-dollar companies.
uncertain
Jev → say → companies
He is now at the vanguard of a wave of low-cost AI systems that, if successful, could rewrite the rules of the booming AI industry.
uncertain
that → rewrite → industry
Almeida, a maths genius raised in Manila in the Philippines, worked at the Google Brain AI lab before joining OpenAI, interrupted by an 18-month sabbatical – much of which he spent playing fantasy video games.
asserted
he → raise → games
At OpenAI, he pioneered the training technique known as reinforcement learning from human feedback, in which bots improve based on how humans rate their answers.
asserted
humans → pioneer → answers
The procedure was used to create ChatGPT, which launched in 2022 as the first AI system able to hold a conversation and answer questions.
asserted
which → use → questions
But chatbots like ChatGPT, which produce results in text, place extraordinary demands on computing resources.
asserted
which → produce → resources
“Language models” such as ChatGPT or Claude produce complex strings of text to reason through questions, and queries can take several seconds or even minutes for a response.
asserted
queries → produce → response
Humans might be willing to wait a few seconds for a chatbot answer.
uncertain
Humans → wait → answer
But increasingly, AI’s users are other machines: a customer service system deciding where to route email complaints; an automated recruitment system screening CVs; or a bank’s compliance software assessing transactions for fraud.
asserted
users → decide → fraud
Jev, released this month, produces answers in nanoseconds.
asserted
Jev → release → nanoseconds
Instead of answering with words, it responds with numbers, for example assigning a probability that a bank payment is fraudulent.
asserted
payment → answer → probability
This makes it orders of magnitude faster and cheaper than rival systems.
asserted
it → make → systems
At a time when corporate customers are seeking to rein in staff spending on AI systems, TypeSafe says Jev delivers results 194 times faster and 445 times cheaper than the best OpenAI and Claude systems.
asserted
Jev → seek → systems
One demo shows the model working fast enough to control the video game Doom.
asserted
model → show → game
In the real world, the concept could be applied to controlling robots fast enough for them to become useful.
uncertain
them → apply → robots
Since the system was released, incessant demand has forced it to suspend sign-ups, and reports have suggested the company’s $US200 million valuation could rise to $US10 billion.
uncertain
valuation → release → billion
Almeida has said investors are “battering down our door”.
asserted
investors → say → door
Vercel, a company that provides software development tools, says Jev became the fastest-adopted AI model in its history after launch.
asserted
Jev → provide → launch
Within days, a cluster of copycat systems has emerged built on the same idea (one, unimaginatively, is called “Kev”).
asserted
one → emerge → idea
But if a swath of AI usage moves from expensive frontier systems to fast, budget models, it could puncture the investment case for OpenAI and Anthropic, both of which are planning to go public in the coming months.
uncertain
both → move → months
HSBC’s analysts said that the system had the potential to deliver a market sell-off similar to the DeepSeek market crash at the start of last year, which sent shares in Nvidia, AMD and Google tumbling on fears that it would lead demand for computing resources to crater.
asserted
demand → say → resources
“The prospect of a new class of AI models that are much cheaper and faster could lead to a resurgence of market fears about AI monetisation,” they said, although they added that those fears would be misplaced.
uncertain
fears → lead → monetisation
Almeida has predicted that his system will lead to a massive expansion in AI usage.
asserted
system → predict → usage
Jev comes from the Jevons paradox, the phenomenon named after the British economist William Stanley Jevons, who observed that when coal use became more efficient, demand for it went up.
asserted
demand → come → it
If the same is true when it comes to AI, then Jev could keep the boom alive.
uncertain
Jev → come → boom
The expensive models it replaces might be collateral damage – Almeida’s earlier creation among them.
uncertain
it → replace → them
The Aporia
