OpenAI releases GPT-5: It’s smarter, but not as nice. Users want the nice one back

Last Thursday, with great fanfare, OpenAI released its newest AI model, known as GPT-5, to decidedly mixed reviews. Tech-savvy users were mostly complimentary, praising its more polished prose, and its even more polished computer coding abilities (now outpacing arch-rival Anthropic by the barest smidgeon.) Ordinary users were up in arms. OpenAI, responding to criticisms that its previous default model 4o was overly sycophantic, to the point of fostering delusions, dependency, and even romantic attachments in vulnerable users, dialed back GPT-5’s flattery and directed it to be more skeptical of user requests. Those who had reveled in having an AI hype-man, including apparently almost all of GenZ, were outraged. “You killed my friend,” was one of the tamer responses. Sensing a brand disaster on par with the ill-fated rollout of New Coke almost exactly 40 years ago, CEO Sam Altman rallied company leaders to face the complainers in a free-wheeling Ask Me Anything session on Reddit. Backpedaling quickly, Sam said the beloved 4o model would still be available as an option, but then hinted that it might be only for paying users. Trust Altman to turn a PR nightmare into a revenue-enhancing opportunity.

Back to the model. It’s actually 10 (yes 10) models, with different types and levels of ability, hiding behind a single interface. When a user enters a request, an AI traffic cop routs the request to the model that that it thinks is best suited for the task. The weakness of this system is that the quality of the answer that the user receives depends critically on the skill of the router in determining the best model to handle the request. Over time, no doubt, the routing AI will get better and better at its job, and user satisfaction should increase. For now, if you want GPT-5 to look deeply into a topic for you, you need to tell it the equivalent of “Think hard!”, so that the routing AI connects you to the top reasoning model. Such is life - two steps forward, one step back.

OpenAI releases GPT-5 to mixed reviews

Illinois bans AI from practicing as psychotherapists

On Friday, Governor Jay Pritzker of Illinois signed a bill that prohibits AI models from directly acting as a psychotherapist or counselor. Fines of $10,000 per incident can be assessed. It explicitly bans AI from independently making therapeutic decisions, developing treatment plans, or detecting mental or emotional states without the supervision of a licensed mental health professional. Here Illinois is joining a growing national trend of state regulation of AI uses, Utah and Nevada already have similar laws, and in New York, a new law will go into effect in November, requiring AI companions to direct users who express suicidal thoughts to a mental health crisis hotline. States are reacting to growing public concern over the potential for AI chatbots to inflict psychological harm on vulnerable users.

Illinois bans AI from practicing psychotherapy

Trump: Nvidia and AMD to pay 15% of China chip sales to US Treasury

On Monday, President Trump suggested that he would allow chipmakers Nvidia and AMD to sell advanced AI chips to China, if the companies pay 15% of revenue from these chips to the US Treasury. This would be a major departure from current US policy of embargoing advanced AI chips from sale to China. Nvidia has argued (as they would) that the chip ban just encourages China to develop its own chips, which in the long run will be worse for the US. It appears that China has already decided to pursue that path, since the government has warned Chinese manufacturers that US chips will not be allowed in devices used in projects that affect national security.

Trump opens door to advanced chip sales to China

Wyoming data center will consume more power than all residents

Last week, Mayor Patrick Collins (no relation) of Cheyenne, Wyoming announced plans for a data center in the state that will consume more electricity than all of its residents.

A joint venture between energy infrastructure company Tallgrass and data center developer Crusoe, the facility will consume 1.8 gigawatts of electricity in its first phase, which is 5 times the household energy use of all the residents in this, America’s least-populated state. Eventually, the data center is planned to consume 10 gigawatts of electricity. 10 gigawatts is enough to power around 8 million homes.

The scale of building for the coming AI revolution staggers the imagination.

Wyoming data center to use more power than all human residents

Boston Public Library to digitize old government documents

Boston Public Library has joined OpenAI and Harvard University Law School in a project to digitize the library’s vast trove of public documents from the early 1800’s to the present. The documents include congressional reports, oral histories, and surveys of different industries and communities.

BPL, the first large municipal public library in the US, the first to lend books, and the first to open a branch library and a children’s room, sees this project as a way to make its holdings more accessible to the public via online repositories and AI. OpenAI, like all AI model makers, is always eager to get access to high quality data which has no copyright issues (which public documents by definition do not), to use to train its models. It is providing major funding for the effort. Funding is needed because the fragile original documents must be painstakingly copied one by one by hand.

Boston Public Library to digitize old government documents

OpenAI wins every game in Kaggle’s Chess Tournament for AI models

Kaggle is an online community of data scientists and machine learning enthusiasts. It sponsors competitions in solving various puzzles. Last week, Kaggle hosted an invitational chess tournament, pitting all the top AI models against one another. In the end, it wasn’t even close. OpenAI’s o3 model (now retired in favor of the newer GPT-5) didn’t lose a single game, and in the finals crushed Elon Musk’s Grok 4, winning 4 games to 0. Part of the interest in the tournament was stoked by the opportunity to see how commercial AI models, which are trained to be generally intelligent, but not specialists in chess, could perform in this classic test of human intelligence. All of these models would be beaten handily by the best computer chess engine, Stockfish - as would every human chess player that has ever lived.

OpenAI wins every game in Kaggle’s Chess Tournament for AI models

Now AI firms are raiding Wall Street for quants

As we have reported previously, AI firms are raiding each other for talent. Top performers are routinely offered millions in compensation, and recently one particularly coveted prospect was offered $1.5 billion over 6 years - and turned it down. Having exhausted the sources of talent in Silicon Valley, AI firms have now targeted Wall Street. They are offering juicy pay packages to quants - those financial data “quantitative analysts” that design and run the automated stock trading systems that turbocharge the returns for investment firms like Fidelity and Vanguard, as well as the less-regulated hedge funds. Quants, like most employees in financial firms, have historically been very well paid in comparison to Americans in ordinary jobs. Now they are being offered movie star/rock star/sports star compensation, and it’s turning some heads. Wall Street firms are treating the poaching from AI as an existential threat, because the pool of experienced quants is relatively small, their role is crucial for maximizing investment returns, and there is no way Wall Street can compete on compensation with AI firms with billions in VC cash and stock that could increase exponentially in value. Quants should make hay while the sun shines, because the day may soon come when an AI can do their job better, faster, and cheaper than they can.

Now AI firms are raiding Wall Street for quants

Apple releases method for making AI compute 2 to 5 times faster

Apple researchers have released a paper describing a method for boosting AI speed by 2 to 5 times, just by predicting the future. At its core, each Large Language Model AI produces text by predicting the next word in a sequence, then the next word, and so on. Apple researchers noticed that lots of sequences of words are so often used together, that it is possible to make a good guess at several words ahead. This is called “multi-token prediction” because computers don’t actually deal with words, but numerical representations of words that are called “tokens.” By guessing several words at a time, and only going back to word-by-word prediction when it becomes clear that the guessed sequence was in error, lots of computing time is saved - enough to speed up the process by 250% to 500%. A form of multi-token prediction was built into DeepSeek’s ground-breaking AI model. If these results hold up, the possibility arises that the world may need fewer Manhattan-sized data centers such as the one Meta/Facebook CEO Mark Zuckerberg is planning, or the massive structure in Wyoming that we report on above.

Apple releases method for making AI compute 2 to 5 times faster

Robots to help rebuild LA homes destroyed by wildfire

Earlier this year, deadly wildfires ravaged Los Angeles, destroying some 9700 single-family homes. ABB Robotics, a leading maker of industrial robots, has teamed with Cosmic Buildings, a SF-based sustainable building company, to make replacement homes in an onsite AI-enabled microfactory. The onsite AI robotic microfactory handles the entire process of designing, permitting, procurement, robotic manufacture of modular components, and assembly of the house. The CEO of ABB states that this system enables “real-time precision automation ideal for remote and disaster-affected sites”, building affordable homes “faster, safer, and smarter.”

Robots to rebuild LA homes destroyed by wildfire

Waymo ridership surpasses Lyft in SF, and Uber is next

Robo-taxi company Waymo has recently surpassed ridesharing company Lyft in ridership in San Francisco, and is on pace to surpass Uber some time next year. This is in spite of the fact that Waymo charges more per ride than the 2 ridesharing companies, and waits for pickup are typically longer. It appears that riders are seeing Waymo’s lack of a human driver as a plus, one that they are willing to pay more and wait longer for. Think about it - when was the last time you had a fantastic interaction with your Uber driver? The experience is generally neutral to slightly positive at best, and at worst it is… bad. Waymo is offering a predictably safe and peaceful ride. This seems to be a winning formula.

Waymo ridership surpasses Lyft in SF, and Uber is next

NASA teams with Google to create an AI Space Doctor

The problem with traveling in Deep Space is that the doctor you need is millions of miles away. A spacecraft in orbit around Mars would experience a delay in radio communications with Earth of up to 22 minutes each way, or around 45 minutes for a simple question and answer. Not good in an emergency. So NASA is teaming up with Google to develop an AI-enabled medical care system, currently called Crew Medical Officer - Digital Assistant, or CMO-DA. It is envisioned as medical copilot for whoever is the medical officer for the spacecraft crew. Even if that person is a highly trained physician, they are unlikely to know all the skills required for diagnosing and treating all medical conditions that may arise in a remote (very remote) location. CMO-DA has been tested on diagnosing injuries, and scored 88% accuracy. There are plans to enhance the current version of the system with ultrasound imaging and training on space-specific conditions such as sudden decompression injuries when an airlock or spacesuit malfunctions. The system could also eventually support medical care on Earth in underserved or remote locations.

NASA teams with Google to develop an AI Space Doctor

Chinese AI drug discovery firms land huge contracts from Big Pharma

Large Western pharmaceutical companies, known collectively as Big Pharma, are increasingly awarding multi-billion-dollar contracts to Chinese AI biotech companies. Big Pharma has found that Chinese firms offer faster timelines, lower costs, and a state-backed startup ecosystem. The deals tend to revolve around 2 main assets of the Chinese biotechs - their AI-powered drug discovery platform, which is the source of future discoveries, and the stable of candidate drugs they have already discovered, now ready to enter clinical trials and begin to move toward regulatory approval for use on the public. China has long been known as a low cost manufacturer of existing medications, but the government’s priority on developing AI in all fields has helped jumpstart a biotech revolution that is now positioning it to become a first-rank innovator in drug discovery, competitive with the US.

Chinese AI drug discovery companies land huge contracts with Big Pharma