OpenAI unveils GPT 4.5 - no “reasoning”, but more “emotional intelligence”

Last Thursday, OpenAI released (to a select few paying users) what it is now calling GPT 4.5. Long code-named “Orion” internally (the source of those little “o” prefixes to o1 and o3), it is the end of an era. From GPT 1 through 2, 3, and the various GPT-4’s, each succeeding model was trained on lots more data, using lots more computer power. And for all previous models, the improvement was obvious, sometimes seemingly magical. Early reviews on the experience with GPT 4.5 is that it’s…. nice, but not thrilling. OpenAI CEO Sam Altman enthused about its “emotional intelligence”, but it is still performing worse on benchmarks in math, science, and coding than smaller models that use the new “reasoning” paradigm that o1 debuted, and DeepSeek used when it shocked the world this past January. Altman has said publicly that 4.5 will be OpenAI’s last “non-reasoning” model. The new paradigm is just that much more powerful. GPT 4.5 is also likely the last model that will be just one model. Altman has hinted that future GPTs will have a variety of models under the hood, hidden behind a user interface that will choose the right model for the user’s query (and pocketbook.)

OpenAI unveils GPT 4.5 - no reasoning, more emotional intelligence

Perplexity powers the first AI-centered smartphone

Deutsche Telekom, T-Mobile’s parent (who knew?) has announced that they will release the first AI-centered smartphone later this year. Instead of tapping apps, the user will talk to an AI assistant (DT calls it a “butler”, how quaint) powered by AI search startup Perplexity. The assistant can “order a taxi, reserve a table, translate in real time” as well as write emails, start phone calls, play music, summarize or translate texts, make calendar entries, and more. In addition to the partnership with Perplexity, DT will have Google Cloud AI, Elevenlabs (speech-to-text and text-to-speech) and Picsart (AI visual design tools) on the AI smartphone. DT is one of the largest telecom companies in the world, and it is making a serious play to become everyone’s favorite AI assistant device.

Perplexity powers the first AI-centered smartphone

Amazon unveils AI Alexa+, but Apple’s AI Siri may be delayed to 2027

These days, the buzz in AI is all about Agents - AI that can actually do stuff in the real world, not just chat about it. Sadly, the original agents, Siri and Alexa, powered by an earlier generation of AI, have not yet retooled with the modern-day, more powerful generative AI. Now Amazon has announced an AI-powered Alexa+, which will have a monthly subscription fee, unless you are a Prime member. Alexa+ will apparently be tightly integrated into both Amazon’s smart home technology and your Amazon account, with the usual chatbot capabilities as well. A phased rollout will begin next month. Meanwhile, the AI upgrade to Apple’s Siri, the OG voice assistant that astonished the world at its 2011 launch, is struggling. It was promised this year, then next year, and now knowledgeable sources say it could be 2027 before it sees the light of day. Two years is an eternity in AI. Perplexity could have sewed up the market by then. (See story above.)

Amazon unveils AI-powered Alexa+

Apple’s AI Siri may not ship until 2027

TSMC pledges to pour $100 billion into US chip making over 4 years

Taiwan Semiconductor Manufacturing Co. (TSMC) is the most advanced computer-chip manufacturing company in the world, and it is responsible for the fabrication of some 90% of all the most powerful AI chips. The Biden Administration thought that Taiwan’s tenuous relationship with nearby China constituted a threat to US national security, and with passage of the CHIPS Act in 2022, encouraged TSMC and other vital suppliers in the semiconductor supply chain to locate facilities in the US. TSMC has already invested some $40 billion in 2 US facilities in Arizona, employing approximately 3,000 people, and has now announced plans to invest an additional $100 billion into US facilities over the next 4 years.

TSMC pledges to pour $100 billion into US chip making over 4 years

AI is causing the fastest transformation of communication in history

The penetration of AI into various categories of business communications has been astonishingly rapid. As noted by Wharton Professor Ethan Mollick, “By September 2024, 18% of financial consumer complaints, 24% of press releases, 15% of job postings & 14% of UN press releases showed signs of LLM writing. And the method undercounts true use.”

AI has caused the most rapid transformation of communication in history

arXiv paper

World’s largest call center uses AI to mask accents of Indian employees

Paris-based outsourcing company Teleperformance, which operates the largest call center in the world, has announced that it is using AI to perform “accent translation” on the speech of its Indian employees who are on calls to English-speaking countries. The company’s deputy CEO crowed about AI’s ability to “neutralize the accent of the Indian speaker with zero latency.” Teleperformance stock fell 11% after the announcement.

World’s largest call center uses AI to mask accents of Indian employees

AI companies follow DeepSeek’s lead to distill smaller, cheaper models

Chinese company DeepSeek shocked the AI world in January when it released an AI model that performed comparably to the best from OpenAI, Google, Anthropic, and Facebook/Meta, even though it claimed to use only a fraction of the computing power needed for the big models. US tech stocks lost a combined $1 trillion in one day, half of that from Nvidia alone, as terrified investors feared that no one would need Nvidia’s ever-pricier AI chips. Cooler heads prevailed, and tech stocks rallied, when it was discovered that DeepSeek had actually piggy-backed off the performance of a larger model, OpenAI’s o1, in a process known as “distilling.” DeepSeek trained their own smaller model with a series of questions and answers posed to OpenAI’s then-leading model, o1. In this way, the larger model, “teaches” the smaller model how to approach and answer queries, and the smaller model gains a great deal of capability without the overhead of a large model. As the saying goes, “Knowing how to think about a problem is worth 20 IQ points.” Now all the big AI companies are starting to distill their large models down to smaller, more efficient models that require a lot less computing power to train and run. But, at least so far, they still have to build the big models in the first place to get the performance that they can then distill down.

AI companies follow DeepSeek’s lead to distill smaller, cheaper models

Waymo providing over 200,000 robotaxi rides/week - 20× 2 years ago

Just 2 years ago, Google’s driverless taxi subsidiary Waymo was only providing 10,000 cab rides a week to its customers. Then by August of last year, it notched a milestone with 100,000 rides per week. Now, barely 6 months later, it is serving 200,000 rides per week in San Francisco, Los Angeles, and Phoenix. Expansions to Austin and Atlanta have been announced for this year, and Miami is slated to open up in 2026.

Waymo serving over 200,000 rides/week - 20x the total 2 years ago

China explores “death dog” robots for urban warfare

It has been credibly reported that scientists associated with the Chinese People’s Liberation Army have carried out tests of robot dogs carrying thermobaric weapons in simulated urban warfare scenarios. Thermobaric weapons, sometimes called “vacuum bombs”, are bombs that spray an ignitable mist into a room and then spark it off, creating a long-lived firestorm so hot that human bodies can be vaporized. Use of thermobaric weapons in a heavily populated urban area may qualify as a war crime, so it seems more of a desperation tactic, reserved for cases of invasion of the Chinese mainland (seems unlikely) or invasion by China of Taiwan (less unlikely, unfortunately.)

China explores “death dog” robots for urban warfare

GOAT is a morphing off-road robot from Switzerland

Switzerland is 70% mountains, so the Swiss are no strangers to rugged terrain. Scientists at Switzerland’s Ecole Polytechnique Federale de Lausanne (EPFL) have developed a morphing robot named GOAT, because it is designed to be Good Over All Terrain. Flexible but durable, the robot can spontaneously shift from a flat “rover” shape to a spindly sphere as needed, so that it can drive, roll, or even swim, all while using less energy than a robot with legs. The design is called “bio-inspired”, because the development team took inspiration from across the animal kingdom, including spiders, kangaroos, cockroaches, and octopuses.

GOAT is a morphing off-road robot from France

Measuring brain aging with MRI and AI

Researchers from the University of Southern California have developed an AI model that can compare a baseline MRI brain scan with follow up brain scans of the same patient to produce a highly accurate estimate of how fast the patient’s brain is aging. This has already been found to be useful in predicting the trajectory of cognitive decline in Alzheimer’s patients.

Measuring brain aging with MRI and AI

PNAS paper

AAFP finds that PCPs are ready to use AI

The American Academy of Family Physicians commissioned a survey of over 1200 primary care physicians last fall, asking about their perceptions of and current use of digital tools in their work, and their interest in AI tools. No surprise, the survey found that the most used digital tool was the electronic medical record, with 87% using it daily and a further 5% using it at least weekly. In contrast, it appears that only third of PCPs have ever used any AI tool in their work to date. Looking forward, over half of respondents felt that AI would improve time to diagnosis, diagnostic accuracy, and appropriateness of treatment. A majority also projected that AI would improve clinician workload, clinician wellbeing, and patient satisfaction. Respondents were less sanguine about the impact of AI on personalization of care, clinician-patient relationships, and clinician job security.

PCPs are ready to use AI