For artificial-intelligence technology and the industry rising up around it in San Francisco and elsewhere, 2024 saw some heady highs.
Two separate Nobel prizes were awarded for AI-related work. AI startups raised unprecedented sums, and some reportedly started to record serious sales numbers. Following early-adopting individuals, companies around the world started to experiment with the technology, with some seeing early success.
Taken together, for enthusiasts, the developments seemed to indicate the technology was well on its way to meeting its promise of being the industry’s biggest leap forward ever.
“I’ve never seen such an incredible amount of activity in a sector or in a technology as quickly as this,” said Rob Siegel, a lecturer in management at Stanford Graduate School of Business who focuses on startups and technology.
At the same time, the shortcomings of the industry and technology — and even the potential danger they pose — also started to become more readily apparent this year.
Reports highlighted the immense environmental impact of the technology and the challenge it poses to the big-tech companies’ promises to cut their carbon emissions. Others exposed how chatbots were sending harmful messages to teens and young women were being tormented by typically teenage boys using sophisticated AI tools to generate faked pornographic images.
Still others brought to light the billions of dollars the biggest AI companies are blowing through even as their technological progress seems to be slowing down and key problems — such as generative AI’s tendency to make up answers whole cloth — remain unresolved.
It became clearer this year that the excitement about the technology and the marketing bluster promoting it has far outpaced its actual capacity to solve real-world problems, said Alex Hanna, the director of research at the Distributed AI Research Institute.
“The hype is huge,” Hanna said. “It’s a hype bubble.”
That it might be, but there were some legitimate reasons for excitement.
Computer scientist Geoffrey Hinton, who studies neural networks used in AI applications, won the 2024 Nobel Prize for physics for foundational work.
Noah Berger/Associated Press File
Through Dec. 13, venture investors had poured $81 billion into U.S.-based artificial-intelligence and machine-learning startups in 2024, according to PitchBook, an industry-research firm. Even though that tally didn’t include the Databricks deal, it was already up 41% year-over-year and nearly at the full-year record of $81.1 billion set in 2021, when the venture industry was booming. More than half of the tally in each of the last two years was invested in San Francisco companies.
While the biggest companies are drawing an outsized share of the capital, venture investors are increasingly spreading money around as different sectors of tech embrace AI. Through the middle of December, AI and machine-learning startups had closed 940 deals nationwide, according to PitchBook. That was up 20% from all of last year. Some 26% of those rounds were closed by San Francisco startups.
Venture investors “are looking for the next wave of untapped use cases and looking for land grabs,” said Brendan Burke, a senior emerging-technology analyst at PitchBook.
Those investments come as growing numbers of companies and investors are experimenting with AI, analysts say. Businesses big and small in different sectors and nations are testing out how they can use the technology in their operations, Siegel said.
Many companies spent the first half of this year trying to figure out how they were going to use AI and which models to build on top of, said Matt Murphy, a partner at Menlo Ventures. Many spent the second half of the year actually deploying the technology and getting it in front of their employees and customers, he said.
“The shift was stark,” Murphy said. “The velocity in which we got there in the second half was impressive.”
FILE – Alphabet CEO Sundar Pichai speaks about Gemini at a Google I/O event in Mountain View, Calif., Tuesday, May 14, 2024. (AP Photo/Jeff Chiu, File)
Jeff Chiu/Associated Press, file
Some companies are already starting to see some successes with using and implementing AI, analysts said. In October, Google CEO Sundar Pichai announced that AI technology was creating about 25% of the new software code for the company’s products.
Siegel said PayPal’s AI-powered chatbot helped him resolve an issue. When he wasn’t paid after selling tickets to an event via the company’s services, he turned to PayPal’s support site, which referred him to its chatbot. The chatbot was able to answer his questions and resolve his issue by offering fast and accurate answers, he said.
“I did not feel like I was talking to a bot,” he said.
Helping it to better perform such tasks and others, AI made some significant technological progress this year, experts say.
One technique embraced by developers such as OpenAI was to have their models take time to consider and evaluate their answers to questions and queries before responding to them. A related technique is “chain of thought,” which involves having models break down questions into their constituent parts and tackling each of those parts individually.
Such techniques can help models deliver better, more accurate results, experts said.
Having the models take more time before answering is kind of like saying, “Hey, use your prefrontal cortex to filter … what you’re going to say and think about it more deeply before you say it,” said James Landay, a professor of computer science at Stanford and co-director of its Human-centered Artificial Intelligence institute.
Unlike the early days of some previous technology booms, companies are already spending real money on AI, Siegel said.
The 950 striking workers will vote Tuesday on ratifying the agreement, according to Unite Here Local 2
Phornthip Korkiatnun said she didn’t envision self as restauranteur. Now, she’s eyeing expansion
The City could receive up to half-inch of rain, in addition to experiencing wind gusts up to 30 miles per hour
“The demand is real and palpable,” he said.
Even so, the industry and technology seemed to lose some of its luster this year.
People are reflected in a window of a hotel at the Davos Promenade in Davos, Switzerland, Jan. 15, 2024.
Markus Schreiber/Associated Press, File
While sales at companies such as Anthropic and OpenAI started to head to the billions of dollars, their losses grew even bigger — with no end in sight. OpenAI reportedly expected to lose at least $5 billion this year. It expects that to swell to $14 billion in 2026, The Information reported in October. Between last year and 2028, it expects to rack up a staggering $44 billion in losses, according to The Information.
AI developers have traditionally created better models by training them on ever-larger sets of data. But each increase in training data generally has required a much greater increase in computing power on needed to process it — typically on pricey, hard-to-come-by, electricity-guzzling AI chips.
The current and projected losses have become so enormous that OpenAI is reportedly considering selling ads to generate additional revenue, Hanna noted. Ads might be a tried-and-true method for internet companies to make money, but it seems “bizarre” that AI companies would have to turn to them to support what was supposed to be something so revolutionary, she said.
The AI developers’ “business model wasn’t viable to begin with,” Hanna said. “Now they’re really finding out how unviable it was.”
Even some AI enthusiasts are questioning the degree to which AI companies will generate a return on the enormous amounts being invested in them. Many are calling the current investment wave a bubble and are expecting a shakeout.
“There will be companies that will fail,” Siegel said.
Industry leaders such as OpenAI might not suffer that fate, but that doesn’t necessarily make them good investments, given how much money they’re raising, he said.
“I don’t know how you get venture-backable returns for the risk that’s being taken,” Siegel said.
But 2024 provided other reasons to question the trajectory of the AI industry and technology besides all the red ink. One big one: technological progress seemed to be slowing down. Much of the development work in AI this year focused on refining existing models rather than creating and releasing newer, bigger and better ones, experts say.
OpenAI released GPT-4, the large-language model underlying ChatGPT, in March 2023. There was some expectation it would release GPT-5 this year, but CEO Sam Altman acknowledged in October that wasn’t likely to happen due to the company focusing on essentially refining its existing models.
A ChapGPT logo is seen in West Chester, Pa., on Dec. 6, 2023.
Matt Rourke/Associated Press, file
A big part of the challenge that companies are facing in developing new models is they’re running out of training data, experts say. They’ve already trained their language models on nearly all the publicly accessible data on the internet. Companies have started to experiment with artificially-generated so-called synthetic data, but there are indications that data isn’t particularly good to train on, and there are fears that such data might actually make the models worse, not better, Landay said.
Indeed, there are indications that progress in model development is plateauing, he and other experts said.
“We haven’t seen the models being released as often as we were seeing a couple years ago,” Landay said. “It’s been slower, and I think the improvements have not been as large.”
Additionally, there was greater focus this year on the shortcomings and dangers of the technology.
Since OpenAI released ChatGPT in late 2022, users have noted its sometimes bizarre or inaccurate responses, what people in the industry soon dubbed “hallucinations.” But in a paper this year, a team of philosophers in Scotland argued that the more appropriate way to describe what AI is doing is “bulls—.” Just like a person who deals in B.S., AI models neither know nor care what is the truth, they argued.
Those models craft their answers to queries by using probabilities, said Jesse Dodge, a senior research scientist at the Allen Institute for AI. They don’t have a concept of what is a true answer; instead, based on their training data, they build their answers one word at a time, determining which word is most likely to come after the previous one.
The probabilistic nature of the model’s responses “means we can’t truly get rid of these hallucinations,” Dodge said.
“The more we use this stuff, the more we’re going to get bad stories of where they go wrong,” Landay said.
The sheer environmental costs of AI technology also came to the fore in 2024. The vast amounts of computing power required to train and run AI models consumes huge and growing amounts of energy. That’s already leading to an uptick in carbon emissions and water consumption, as water is evaporated to cool the computers running AI models and the power plants providing their electricity.
Projections made this year indicate those problems are only going to get worse as more and more data centers are built to power the AI boom. Utilities are already keeping online coal-fired plants that were scheduled to close and are gearing up to build natural-gas-fired plants to meet the soaring demand for energy from AI developers. And companies such as Microsoft and Google acknowledged that they are falling short of their pledges to eliminate their carbon emissions.
“These models are increasingly expensive, and emissions from the large tech companies are only going up,” Dodge said.
If you have a tip about tech, startups or the venture industry, contact Troy Wolverton at [email protected] or via text or Signal at 415.515.5594.
Copyright for syndicated content belongs to the linked Source link
