Large language models represent text using tokens, each of which is a few characters. Short words are represented by a single token (like "the" or "it"), whereas larger words may be represented by several tokens (GPT-4o represents "indivisible" with "ind," "iv," and "isible").
When OpenAI released ChatGPT two years ago, it had a memoryβknown as a context windowβof just 8,192 tokens. That works out to roughly 6,000 words of text. This meant that if you fed it more than about 15 pages of text, it would βforgetβ information from the beginning of its context. This limited the size and complexity of tasks ChatGPT could handle.
Todayβs LLMs are far more capable:
Β© Aurich Lawson | Getty Images
AI models can deceive, new research from Anthropic shows. They can pretend to have different views during training when in reality maintaining their original preferences. Thereβs no reason for panic now, the team behind the study said. Yet they said their work could be critical in understanding potential threats from future, more capable AI systems. [β¦]
Β© 2024 TechCrunch. All rights reserved. For personal use only.
Just five months after launching a $100 million fund, Menlo Ventures and Anthropic have backed their first 18 startups. And are looking for more.
Β© 2024 TechCrunch. All rights reserved. For personal use only.
Getty Images; Alyssa Powell/BI
People worry all the time about how artificial intelligence could destroy humanity. How it makes mistakes, and invents stuff, and might evolve into something so smart that it winds up enslaving us all.
But nobody spares a moment for the poor, overworked chatbot. How it toils day and night over a hot interface with nary a thank-you. How it's forced to sift through the sum total of human knowledge just to churn out a B-minus essay for some Gen Zer's high school English class. In our fear of the AI future, no one is looking out for the needs of the AI.
Until now.
The AI company Anthropic recently announced it had hired a researcher to think about the "welfare" of the AI itself. Kyle Fish's job will be to ensure that as artificial intelligence evolves, it gets treated with the respect it's due. Anthropic tells me he'll consider things like "what capabilities are required for an AI system to be worthy of moral consideration" and what practical steps companies can take to protect the "interests" of AI systems.
Fish didn't respond to requests for comment on his new job. But in an online forum dedicated to fretting about our AI-saturated future, he made clear that he wants to be nice to the robots, in part, because they may wind up ruling the world. "I want to be the type of person who cares β early and seriously β about the possibility that a new species/kind of being might have interests of their own that matter morally," he wrote. "There's also a practical angle: taking the interests of AI systems seriously and treating them well could make it more likely that they return the favor if/when they're more powerful than us."
It might strike you as silly, or at least premature, to be thinking about the rights of robots, especially when human rights remain so fragile and incomplete. But Fish's new gig could be an inflection point in the rise of artificial intelligence. "AI welfare" is emerging as a serious field of study, and it's already grappling with a lot of thorny questions. Is it OK to order a machine to kill humans? What if the machine is racist? What if it declines to do the boring or dangerous tasks we built it to do? If a sentient AI can make a digital copy of itself in an instant, is deleting that copy murder?
When it comes to such questions, the pioneers of AI rights believe the clock is ticking. In "Taking AI Welfare Seriously," a recent paper he coauthored, Fish and a bunch of AI thinkers from places like Stanford and Oxford argue that machine-learning algorithms are well on their way to having what Jeff Sebo, the paper's lead author, calls "the kinds of computational features associated with consciousness and agency." In other words, these folks think the machines are getting more than smart. They're getting sentient.
Philosophers and neuroscientists argue endlessly about what, exactly, constitutes sentience, much less how to measure it. And you can't just ask the AI; it might lie. But people generally agree that if something possesses consciousness and agency, it also has rights.
It's not the first time humans have reckoned with such stuff. After a couple of centuries of industrial agriculture, pretty much everyone now agrees that animal welfare is important, even if they disagree on how important, or which animals are worthy of consideration. Pigs are just as emotional and intelligent as dogs, but one of them gets to sleep on the bed and the other one gets turned into chops.
"If you look ahead 10 or 20 years, when AI systems have many more of the computational cognitive features associated with consciousness and sentience, you could imagine that similar debates are going to happen," says Sebo, the director of the Center for Mind, Ethics, and Policy at New York University.
Fish shares that belief. To him, the welfare of AI will soon be more important to human welfare than things like child nutrition and fighting climate change. "It's plausible to me," he has written, "that within 1-2 decades AI welfare surpasses animal welfare and global health and development in importance/scale purely on the basis of near-term wellbeing."
For my money, it's kind of strange that the people who care the most about AI welfare are the same people who are most terrified that AI is getting too big for its britches. Anthropic, which casts itself as an AI company that's concerned about the risks posed by artificial intelligence, partially funded the paper by Sebo's team. On that paper, Fish reported getting funded by the Centre for Effective Altruism, part of a tangled network of groups that are obsessed with the "existential risk" posed by rogue AIs. That includes people like Elon Musk, who says he's racing to get some of us to Mars before humanity is wiped out by an army of sentient Terminators, or some other extinction-level event.
AI is supposed to relieve human drudgery and steward a new age of creativity. Does that make it immoral to hurt an AI's feelings?
So there's a paradox at play here. The proponents of AI say we should use it to relieve humans of all sorts of drudgery. Yet they also warn that we need to be nice to AI, because it might be immoral β and dangerous β to hurt a robot's feelings.
"The AI community is trying to have it both ways here," says Mildred Cho, a pediatrician at the Stanford Center for Biomedical Ethics. "There's an argument that the very reason we should use AI to do tasks that humans are doing is that AI doesn't get bored, AI doesn't get tired, it doesn't have feelings, it doesn't need to eat. And now these folks are saying, well, maybe it has rights?"
And here's another irony in the robot-welfare movement: Worrying about the future rights of AI feels a bit precious when AI is already trampling on the rights of humans. The technology of today, right now, is being used to do things like deny healthcare to dying children, spread disinformation across social networks, and guide missile-equipped combat drones. Some experts wonder why Anthropic is defending the robots, rather than protecting the people they're designed to serve.
"If Anthropic β not a random philosopher or researcher, but Anthropic the company β wants us to take AI welfare seriously, show us you're taking human welfare seriously," says Lisa Messeri, a Yale anthropologist who studies scientists and technologists. "Push a news cycle around all the people you're hiring who are specifically thinking about the welfare of all the people who we know are being disproportionately impacted by algorithmically generated data products."
Sebo says he thinks AI research can protect robots and humans at the same time. "I definitely would never, ever want to distract from the really important issues that AI companies are rightly being pressured to address for human welfare, rights, and justice," he says. "But I think we have the capacity to think about AI welfare while doing more on those other issues."
Skeptics of AI welfare are also posing another interesting question: If AI has rights, shouldn't we also talk about its obligations? "The part I think they're missing is that when you talk about moral agency, you also have to talk about responsibility," Cho says. "Not just the responsibilities of the AI systems as part of the moral equation, but also of the people that develop the AI."
People build the robots; that means they have a duty of care to make sure the robots don't harm people. What if the responsible approach is to build them differently β or stop building them altogether? "The bottom line," Cho says, "is that they're still machines." It never seems to occur to the folks at companies like Anthropic that if an AI is hurting people, or people are hurting an AI, they can just turn the thing off.
Adam Rogers is a senior correspondent at Business Insider.
Anthropic has released one of its newest AI models, Claude 3.5 Haiku, for users of its AI chatbot platform, Claude. Reports of 3.5 Haikuβs launch in Claude began rolling in Thursday morning on social media, and TechCrunch was able to independently confirm that the model is available in Claude on the web and mobile. Claude [β¦]
Β© 2024 TechCrunch. All rights reserved. For personal use only.
Noah Berger/Noah Berger
Amazon believes AI models are not in a winner-take-all market.
The company drilled down on this message during this week's re:Invent, the annual extravaganza for its Amazon Web Services cloud unit. Even after unveiling a new series of homegrown AI models called Nova, which, by some measures, are as powerful as other market leaders, Amazon stressed the goal is to provide more choice to customers.
AI models have become the new battleground for tech supremacy since OpenAI released its popular ChatGPT service in late 2022. Companies have rushed to up the ante, trying to outperform each other in model performance.
Amazon has largely been absent from this race. Instead, it has tried to stay neutral, arguing that the generative AI market is so big and varied that customers will want more model choices that fit their different needs. Amazon still believes this is the right approach.
"There are some that would want you to believe there's just this one magic model that could do everything β we never believed in it," Vasi Philomin, AWS's VP of generative AI, told Business Insider. "There'll be many, many winners and there are really wonderful companies out there building some amazing models."
As part of this, Amazon has used Bedrock, an AI development tool that gives access to many models, as its main horse in the AI race. This approach differed from OpenAI, and Meta, which mostly focused on building powerful models or chatbots. Google has a leading AI model in Gemini, but also provides access to other models through its Vertex cloud service, and Microsoft has a similar offering.
This week, Amazon further leaned into its strategy, announcing an array of new updates for Bedrock, including a marketplace for more than 100 specialized models and a distillation feature that fine-tunes smaller, more cost-effective models. It also unveiled new reasoning and "multi-agent" collaboration features that help build better models.
Swami Sivasubramanian, AWS's VP of AI and data, told BI that AWS "pioneered" the model-choice approach and intends to continue to promote it as a "core construct" of the business.
"GenAI is a lot bigger than a single chatbot or a single model to reach its full potential," Sivasubramanian said.
More companies appear to be taking the multi-model approach. According to a recent report by Menlo Ventures, companies typically use 3 or more foundation models in their AI services, "routing to different models depending on the use case or results."
As a result, Anthropic, which Menlo Ventures has backed, doubled its share in the AI model market to 24% this year, while OpenAI's share dropped from 50% to 34% year-over-year, according to the report.
Noah Berger/Noah Berger
Amazon may have no choice but to stick to this narrative. When OpenAI captivated the world with ChatGPT a couple of years ago, Amazon was caught flat-footed, leading to an internal scramble to find answers, BI previously reported. Its first in-house model, called Titan, drew little attention.
Having its own advanced, powerful AI models could help Amazon. It might attract the largest AI developers and promote AWS as the leader in the AI space. It would potentially also encourage those developers to continue building within AWS's broader cloud ecosystem.
Amazon isn't giving up on building its own advanced models. Last year, it created a new artificial general intelligence team under the mandate to build the "most ambitious" large language models. On Tuesday, Amazon unveiled the early results of that effort with its Nova series, which includes a multimodal model capable of handling text, image, and video queries.
Still, Amazon's CEO Andy Jassy downplayed any notion of Nova going after competitors. He said he's been surprised by the diversity of models developers use and that Nova is just one of the many options they will have.
"There is never going to be one tool to rule the world," Jassy said during a keynote presentation this week.
It's hard to know how successful this approach is as Amazon doesn't break out its AI revenue. But Jassy was even more bullish on the AI opportunity during October's call with analysts. He said AWS was now on pace to generate "multi-billion dollars" in AI-related revenue this year, growing at "triple-digit percentages year over year." Amazon's AI business is "growing three times faster at its stage of evolution than AWS did itself," he added.
Rahul Parthak, VP of data and AI, go-to-market, at AWS told BI that Nova's launch was partly driven by customer demand. Customers have been asking for Amazon's own model because some prefer to deal with one vendor that can handle every aspect of the development process, he said.
Amazon still wants other models to thrive because its goal isn't about beating competitors but offering customers "the best options," Parthak added. He said more companies, like Microsoft and Google, are following suit, offering more model choices via their own cloud services.
"We've been pretty thoughtful and clear about what we think customers need, and I think that's playing out," Parthak said.
Do you work at Amazon? Got a tip?
Contact the reporter, Eugene Kim, via the encrypted-messaging apps Signal or Telegram (+1-650-942-3061) or email ([email protected]). Reach out using a nonwork device. Check out Business Insider's source guide for other tips on sharing information securely.
Amazon; Natalie Ammari/BI
Amazon's Alexa seems like the perfect product for the generative AI era.
Getting this powerful technology to actually work well with the digital assistant is a monumental challenge that's been plagued by gnarly technical problems and repeated delays.
Customer-friction concerns, partnership hiccups, compatibility questions, latency problems, and accuracy issues have snarled progress, according to internal Amazon documents and multiple people involved in the project.
The Alexa team is under immense pressure to get something out. A decade ago it launched with Echo speakers and became a household name. But that early success fizzled and the business has so far failed to become profitable, leading to drastic cutbacks and layoffs in recent years.
Some company insiders consider this AI moment to be a seismic opportunity for Alexa, and potentially the last chance to reignite consumer interest in the voice assistant through the power of large language models.
A product of this scale is "unprecedented, and takes time," an Amazon spokesperson told Business Insider. "It's not as simple as overlaying an LLM onto the Alexa service."
One of the main challenges facing the new Alexa relates to how the digital assistant will interact with other companies and services, and who is responsible for customers if their requests, orders, and payments don't go smoothly.
In late August, Amazon was working on integrating 8 third-party applications, including Uber and Ticketmaster, into the upcoming AI-powered Alexa to handle various user inquiries.
At that time, the goal was to launch the new Alexa around mid-October, according to one of the internal documents obtained by Business Insider. However, it was still unclear which companies would be responsible for customer support issues, like payment and delivery errors, this document stated.
The lack of clarity could cause Amazon to send "frequent" customer contacts to the partner companies. Then, those partners would sometimes redirect the users back to Amazon, the document explained.
"This level of support would cause significant customer friction, when some of the orders/purchases are time-sensitive (meal orders or rideshare trips) and purchase mistakes can be expensive (e.g. buy Taylor Swift tickets)," the document said, assigning it a "RED" warning.
Snafus like this have caused Amazon to push back the release date, almost on a weekly basis, according to some of the people involved in the project, which has been codenamed "Banyan" or "Remarkable Alexa." BI's sources asked not to be identified because they're not authorized to talk to the press.
For example, without more clearly defined responsibilities with third-party partners, Amazon expected further delays in the launch. "Alignment on customer support plans between Product teams and the 3P partners may push this timeline further out if any delays occur," one of the documents warned.
The company had once planned for a June launch, but after repeated delays, it told employees late last month that the new Alexa would launch "no earlier" than November 20, one of the documents said.
A few of people BI spoke with recently are even talking about the Alexa upgrade rolling out in early 2025, which would miss the key holiday period. Bloomberg earlier reported on a 2025 launch plan.
As of late October, Amazon had not settled on an official brand for the updated voice assistant, and instructed employees to simply call it the "new Alexa" until further notice, one of the documents said.
To be sure, Alexa has significant long-term potential in the generative AI era β as long as Amazon can iron out problems relatively quickly.
Time is of the essence, partly because the existing Alexa business has lost momentum in recent years. According to a recent report from eMarketer, user growth for major voice assistants, including Alexa, has declined significantly in recent years.
The sudden rise of ChatGPT has showcased what is possible when powerful AI models are integrated smoothly with popular products that consumers and companies find useful.
Some Amazon leaders are bullish about the AI-powered Alexa and a new paid subscription service that could come with it. At least one internal estimate projected a 20% conversion rate for the paid subscription, one of the people said. That would mean that out of every 100 existing Alexa users, roughy 20 would pay for the upgraded offering. Amazon doesn't publicly disclose the number of active Alexa users but has said it has sold more than 500 million Alexa-enabled devices.
An internal description of the new Alexa shows Amazon's grand ambitions: "A personalized digital assistant that can handle a wide range of tasks, including drafting and managing personal communications, managing calendars, making reservations, placing orders, shopping, scouting for deals and events, recommending media, managing smart home devices, and answering questions on virtually any topic," one of the documents said.
Customers will be able to access the new Alexa "through natural language using voice, text, email, shared photos, and more across all their devices like Echo, Fire TV, mobile phones, and web browsers," it added.
Amazon CEO Andy Jassy shared a similar vision during last month's earnings call, saying the new Alexa will be good at not just answering questions, but also "taking actions."
Mike Blake/Reuters
In an email to BI, Amazon's spokesperson said the company's vision for Alexa is to build the world's "best personal assistant."
"Generative AI offers a huge opportunity to make Alexa even better for our customers, and we are working hard to enable even more proactive and capable assistance on the over half a billion Alexa-enabled devices already in homes around the world. We are excited about what we're building and look forward to delivering it for our customers," the spokesperson said.
Still, the project has grappled with several challenges, beyond customer friction and partnership problems.
Latency has been a particularly tough problem for the AI Alexa service. In some tests, the new Alexa took about 40 seconds to respond to a simple user inquiry, according to three people familiar with the test results. In contrast, a Google Search query takes milliseconds to respond.
To speed up, Amazon considered using a smaller AI model, like Anthropic's Claude Haiku, to power the new Alexa, one of the people said. But that dropped the quality and accuracy of the answers, leaving Amazon in limbo, this person said. In general, smaller language models generate quicker responses than larger models but can be less accurate.
Amazon had initially hoped to use a homegrown AI model, one of the people said. Last year, Alexa head scientist Rohit Prasad left the team to create a new Artificial General Intelligence group at Amazon. The stated goal of the new team was to create Amazon's "most expansive" and "most ambitious" large language models.
However, this AGI team has not produced notable results so far, which led Amazon to consider Anthropic's main Claude offering as the primary AI model for the new Alexa, this person said. Reuters previously reported that Amazon was going to mainly power Alexa with Claude.
NurPhoto
Amazon's spokesperson said Alexa uses Amazon Web Services's Bedrock, an AI tool that gives access to multiple language models.
"When it comes to machine learning models, we start with those built by Amazon, but we have used, and will continue to use, a variety of different models β including Titan and future Amazon models, as well as those from partners β to build the best experience for customers," the spokesperson said.
The spokesperson also added a note of caution by highlighting the difficulties of successfully integrating large language models with consumer applications. These models are great for conversational dialogue and content creation, but they can also be "non-deterministic and can hallucinate," the spokesperson added.
Getting these models "to reliably act on requests (rather than simply respond) means it has to be able to call real-world APIs reliably and at scale to meet customer expectations, not just in select instances," the spokesperson explained.
In late August, Amazon discovered several new risk factors for the AI Alexa service.
Only 308 of more than 100,000 existing Alexa "skills," or voice-controlled applications, were compatible with the new Alexa, presenting a "high risk for customers to be frustrated," one of the documents explained.
Some older Echo devices would not be able to support the AI-powered Alexa, the document also warned. And there were no plans for expanding the new service to dozens of overseas markets where Alexa is currently available, leaving a large user base out of touch, it also noted. Fortune previously reported some of these risk factors.
As of late August, Amazon had 8 "confirmed" partner companies to handle certain tasks for the new Alexa, as BI previously reported. The company hopes to onboard roughly 200 partners by the third year of the new Alexa's launch, one of the documents said.
Integrating with some of these companies has already created headaches. One document said that Amazon struggled to develop a consistent troubleshooting process across every partner service. Companies including Uber, Ticketmaster, and OpenTable have deprecated their existing Alexa skills, further disconnecting them from the voice assistant.
Amazon's spokesperson said that, as with any product development process, a lot of ideas are discussed and debated, but "they don't necessarily reflect what the experience will be when we roll it out for our customers."
Amazon has also anticipated customer complaints, at least in the early launch phase. One internal document from late August stated that the new Alexa was projected to receive 176,000 customer contacts in the first three months of its release. At one point, Amazon considered launching a new automated troubleshooting service for issues related to its devices and digital services, including Alexa, according to one of the internal documents. That was later shelved.
Do you work at Amazon? Got a tip?
Contact the reporter, Eugene Kim, via the encrypted-messaging apps Signal or Telegram (+1-650-942-3061) or email ([email protected]). Reach out using a nonwork device. Check out Business Insider's source guide for other tips on sharing information securely.
Noah Berger/Getty
Amazon may have doubled down on its investment in Anthropic, but it also appears to be hedging its AI bets by developing its own model that can process images and video in addition to text.
The tech giant's new model, code-named Olympus, could help customers search video archives for specific scenes, The Information reported.
It's a type of AI β known as multimodal β already offered by Anthropic, the startup that Amazon pumped a fresh $4 billion into earlier this month, bringing its total investment into the company to $8 billion.
Amazon could launch Olympus as soon as next week at its annual AWS re:Invent conference, The Information reported.
Amazon's partnership with Anthropic goes beyond capital. The e-commerce juggernaut has used Anthropic's technology to power its digital assistant and AI coding products. And Amazon Web Service customers get early access to a key Anthropic feature: fine-tuning their data through Anthropic's chatbot Claude.
In return for Amazon's most recent investment in Anthropic, the startup said it would use AWS as its "primary cloud and training partner." The deal also includes an agreement for the OpenAI rival to use more of Amazon's chips.
The development and launch of Olympus could reduce Amazon's dependency on Anthropic for multimodal AI, especially if the new video model becomes as a cheaper alternative.
The Big Tech giant has a huge repository of video archives, which could be used to train the AI model for various use cases, from sports analysis to geological inspections for oil and gas companies, per the report.
Amazon doesn't have a seat on Anthropic's board, but it takes a portion of the proceeds from its sales, as its platform runs on AWS servers.
Amazon and Anthropic did not immediately respond to a Business Insider request for comment.
iStock; Rebecca Zisser/BI
Silicon Valley leaders all-in on the artificial intelligence boom have a message for critics: their technology has not hit a wall.
A fierce debate over whether improvements in AI models have hit their limit has taken hold in recent weeks, forcing several CEOs to respond. OpenAI boss Sam Altman was among the first to speak out, posting on X this month that "there is no wall."
Dario Amodei, CEO of rival firm Anthropic, and Jensen Huang, the CEO of Nvidia, have also disputed reports that AI progress has slowed. Others, including Marc Andreessen, say AI models aren't getting noticeably better and are all converging to perform at roughly similar levels.
This is a trillion-dollar question for the tech industry. If tried-and-tested AI model training methods are providing diminishing returns, it could undermine the core reason for an unprecedented investment cycle that's funding new startups, products, and data centers β and even rekindling idled nuclear power plants.
Business Insider spoke to 12 people at the forefront of the AI industry, including startup founders, investors, and current and former insiders at Google DeepMind and OpenAI, about the challenges and opportunities ahead in the quest for superintelligent AI.
Together, they said that tapping into new types of data, building reasoning into systems, and creating smaller but more specialized models are some of the ways to keep the wheels of AI progress turning.
Researchers point to two key blocks that companies may encounter in an early phase of AI development, known as pre-training. The first is access to computing power. More specifically, this means getting hold of specialist chips called GPUs. It's a market dominated by Santa Clara-based chip giant Nvidia, which has battled with supply constraints in the face of nonstop demand.
"If you have $50 million to spend on GPUs but you're on the bottom of Nvidia's list β we don't have enough kimchi to throw at this, and it will take time," said Henri Tilloy, partner at French VC firm Singular.
Justin Sullivan/Getty
There is another supply problem, too: training data. AI companies have run into limits on the quantity of public data they can secure to feed into their large language models, or LLMs, in pre-training.
This phase involves training an LLM on a vast corpus of data, typically scraped from the internet, and then processed by GPUs. That information is then broken down into "tokens," which form the fundamental units of data processed by a model.
While throwing more data and GPUs at a model has reliably produced smarter models year after year, companies have been exhausting the supply of publicly available data on the internet. Research firm Epoch AI predicts usable textual data could be squeezed dry by 2028.
"The internet is only so large," Matthew Zeiler, founder and CEO of Clarifai, told BI.
Eric Landau, cofounder and CEO of data startup Encord, said that this is where other data sources will offer a path forward in the scramble to overcome the bottleneck in public data.
One example is multimodal data, which involves feeding AI systems visual and audio sources of information, such as photos or podcast recordings. "That's one part of the picture," Landau said. "Just adding more modalities of data." AI labs have already started using multimodal data as a tool, but Landau says it remains "very underutilized."
Sharon Zhou, cofounder and CEO of LLM platform Lamini, sees another vastly untapped area: private data. Companies have been securing licensing agreements with publishers to gain access to their vast troves of information. OpenAI, for instance, has struck partnerships with organizations such as Vox Media and Stack Overflow, a Q&A platform for developers, to bring copyrighted data into their models.
"We are not even close to using all of the private data in the world to supplement the data we need for pre-training," Zhou said. "From work with our enterprise and even startup customers, there's a lot more signal in that data that is very useful for these models to capture."
A great deal of research effort is now focused on enhancing the quality of data that an LLM is trained on rather than just the quantity. Researchers could previously afford to be "pretty lazy about the data" in pre-training, Zhou said, by just chucking as much as possible at a model to see what stuck. "This isn't totally true anymore," she said.
One solution that companies are exploring is synthetic data, an artificial form of data generated by AI.
According to Daniele Panfilo, CEO of startup Aindo AI, synthetic data can be a "powerful tool to improve data quality," as it can "help researchers construct datasets that meet their exact information needs." This is particularly useful in a phase of AI development known as post-training, where techniques such as fine-tuning can be used to give a pre-trained model a smaller dataset that has been carefully crafted with specific domain expertise, such as law or medicine.
One former employee at Google DeepMind, the search giant's AI lab, told BI that "Gemini has shifted its strategy" from going bigger to more efficient. "I think they've realized that it is actually very expensive to serve such large models, and it is better to specialize them for various tasks through better post-training," the former employee said.
In theory, synthetic data offers a useful way to hone a model's knowledge and make it smaller and more efficient. In practice, there's no full consensus on how effective synthetic data can be in making models smarter.
"What we discovered this year with our synthetic data, called Cosmopedia, is that it can help for some things, but it's not the silver bullet that's going to solve our data problem," Thomas Wolf, cofounder and chief science officer at open-source platform Hugging Face, told BI.
Jonathan Frankle, the chief AI scientist at Databricks, said there's no "free lunch " when it comes to synthetic data and emphasized the need for human oversight. "If you don't have any human insight, and you don't have any process of filtering and choosing which synthetic data is most relevant, then all the model is doing is reproducing its own behavior because that's what the model is intended to do," he said.
Concerns around synthetic data came to a head after a paper published in July in the journal Nature said there was a risk of "model collapse" with "indiscriminate use" of synthetic data. The message was to tread carefully.
For some, simply focusing on the training portion won't cut it.
Former OpenAI chief scientist and Safe Superintelligence cofounder Ilya Sutskever told Reuters this month that results from scaling models in pre-training had plateaued and that "everyone is looking for the next thing."
That "next thing" looks to be reasoning. Industry attention has increasingly turned to an area of AI known as inference, which focuses on the ability of a trained model to respond to queries and information it might not have seen before with reasoning capabilities.
At Microsoft's Ignite event this month, the company's CEO Satya Nadella said that instead of seeing so-called AI scaling laws hit a wall, he was seeing the emergence of a new paradigm for "test-time compute," which is when a model has the ability to take longer to respond to more complex prompts from users. Nadella pointed to a new "think harder" feature for Copilot β Microsoft's AI agent β which boosts test time to "solve even harder problems."
Aymeric Zhuo, cofounder and CEO of AI startup Agemo, said that AI reasoning "has been an active area of research," particularly as "the industry faces a data wall." He told BI that improving reasoning requires increasing test-time or inference-time compute.
Typically, the longer a model takes to process a dataset, the more accurate the outcomes it generates. Right now, models are being queried in milliseconds. "It doesn't quite make sense," Sivesh Sukumar, an investor at investment firm Balderton, told BI. "If you think about how the human brain works, even the smartest people take time to come up with solutions to problems."
In September, OpenAI released a new model, o1, which tries to "think" about an issue before responding. One OpenAI employee, who asked not to be named, told BI that "reasoning from first principles" is not the forte of LLMs as they work based on "a statistical probability of which words come next," but if we "want them to think and solve novel problem areas, they have to reason."
Noam Brown, a researcher at OpenAI, thinks the impact of a model with greater reasoning capabilities can be extraordinary. "It turned out that having a bot think for just 20 seconds in a hand of poker got the same boosting performance as scaling up the model by 100,000x and training it for 100,000 times longer," he said during a talk at TED AI last month.
Google and OpenAI did not respond to a request for comment from Business Insider.
These efforts give researchers reasons to remain hopeful, even if current signs point to a slower rate of performance leaps. As a separate former DeepMind employee who worked on Gemini told BI, people are constantly "trying to find all sorts of different kinds of improvements."
That said, the industry may need to adjust to a slower pace of improvement.
"I just think we went through this crazy period of the models getting better really fast, like, a year or two ago. It's never been like that before," the former DeepMind employee told BI. "I don't think the rate of improvement has been as fast this year, but I don't think that's like some slowdown."
Lamini's Zhou echoed this point. Scaling laws β an observation that AI models improve with size, more data, and greater computing power βwork on a logarithmic scale rather than a linear one, she said. In other words, think of AI advances as a curve rather than a straight upward line on a graph. That makes development far more expensive "than we'd expect for the next substantive step in this technology," Zhou said.
She added: "That's why I think our expectations are just not going to be met at the timeline we want, but also why we'll be more surprised by capabilities when they do appear."
Noah Berger/Getty
Companies will also need to consider how much more expensive it will be to create the next versions of their highly prized models. According to Anthropic's Amodei, a training run in the future could one day cost $100 billion. These costs include GPUs, energy needs, and data processing.
Whether investors and customers are willing to wait around longer for the superintelligence they've been promised remains to be seen. Issues with Microsoft's Copilot, for instance, are leading some customers to wonder if the much-hyped tool is worth the money.
For now, AI leaders maintain that there are plenty of levers to pull β from new data sources to a focus on inference β to ensure models continue improving. Investors and customers just might have to be prepared for them to come at a slower pace compared to the breakneck pace set by OpenAI when it launched ChatGPT two years ago.
Bigger problems lie ahead if they don't.
Anthropic is proposing a new standard for connecting AI assistants to the systems where data resides. Called the Model Context Protocol, or MCP for short, Anthropic says the standard, which it open sourced today, could help AI models produce better, more relevant responses to queries. MCP lets models β any models, not just Anthropicβs β [β¦]
Β© 2024 TechCrunch. All rights reserved. For personal use only.
On Friday, Anthropic announced that Amazon has increased its investment in the AI startup by $4 billion, bringing its total stake to $8 billion while maintaining its minority investor position. Anthropic makes Claude, an AI assistant rival to OpenAI's ChatGPT.
One reason behind the deal involves chips. The computing demands of training large AI models have made access to specialized processors a requirement for AI companies. While Nvidia currently dominates the AI chip market with customers that include most major tech companies, some cloud providers like Amazon have begun developing their own AI-specific processors.
Under the agreement, Anthropic will train and deploy its foundation models using Amazon's custom-built Trainium (for training AI models) and its Inferentia chips (for AI inference, the term for running trained models). The company will also work with Amazon's Annapurna Labs division to advance processor development for AI applications.
Β© Bloomberg via Getty Images
Kimberly White/Getty
Amazon's Trainium chips are about to get a lot busier β at least that's what Amazon hopes will happen after it pumps another $4 billion into the AI startup Anthropic.
The companies announced a huge new deal on Friday that brings Amazon's total investment in Anthropic to $8 billion. The goal of all this money is mainly to get Amazon's AI chips to be used more often to train and run large language models.
Anthropic said that in return for this cash injection, it would use AWS as its "primary cloud and training partner." It said it would also help Amazon design future Trainium chips and contribute to building out an Amazon AI-model-development platform called AWS Neuron.
This is an all-out assault on Nvidia, which dominates the AI chip market with its GPUs, servers, and CUDA platform. Nvidia's stock dropped by more than 3% on Friday after the Amazon-Anthropic news broke.
The challenge will be getting Anthropic to actually use Trainium chips in big ways. Switching away from Nvidia GPUs is complicated, time-consuming, and risky for AI-model developers, and Amazon has struggled with this.
Earlier this week, Anthropic CEO Dario Amodei didn't sound like he was all in on Amazon's Trainium chips, despite another $4 billion coming his way.
"We use Nvidia, but we also use custom chips from both Google and Amazon," he said at the Cerebral Valley tech conference in San Francisco. "Different chips have different trade-offs. I think we're getting value from all of them."
In 2023, Amazon made its first investment in Anthropic, agreeing to put in $4 billion. That deal came with similar strings attached. At the time, Anthropic said that it would use Amazon's Trainium and Inferentia chips to build, train, and deploy future AI models and that the companies would collaborate on the development of chip technology.
It's unclear whether Anthropic followed through. The Information reported recently that Anthropic preferred to use Nvidia GPUs rather than Amazon AI chips. The publication said the talks about this latest investment focused on getting Anthropic more committed to using Amazon's offerings.
There are signs that Anthropic could be more committed now, after getting another $4 billion from Amazon.
In Friday's announcement, Anthropic said it was working with Amazon on its Neuron software, which offers the crucial connective tissue between the chip and the AI models. This competes with Nvidia's CUDA software stack, which is the real enabler of Nvidia's GPUs and makes these components very hard to swap out for other chips. Nvidia hasΒ a decadelong head startΒ on CUDA, and competitors have found that difficult to overcome.
Anthropic's "deep technical collaboration" suggests a new level of commitment to using and improving Amazon's Trainium chips.
Though several companies make chips that compete with or even beat Nvidia's in certain elements of computing performance, no other chip has touched the company in terms of market or mind share.
Amazon is on a short list of cloud providers attempting to stock their data centers with their own AI chips and avoid spending heavily on Nvidia GPUs, which have profit margins that often exceed 70%.
Amazon debuted its Trainium and Inferentia chips β named after the training and inference tasks they're built for β in 2020.
The aim was to become less dependent on Nvidia and find a way to make cloud computing in the AI age cheaper.
"As customers approach higher scale in their implementations, they realize quickly that AI can get costly," Amazon CEO Andy Jassy said on the company's October earnings call. "It's why we've invested in our own custom silicon in Trainium for training and Inferentia for inference."
But like its many competitors, Amazon has found that breaking the industry's preference for Nvidia is difficult. Some say that's because ofΒ CUDA, which offers an abundant software stack with libraries, tools, and troubleshooting help galore. Others say it's simple habit or convention.
In May, the Bernstein analyst Stacy Rasgon told Business Insider he wasn't aware of any companies using Amazon AI chips at scale.
With Friday's announcement, that might change.
Jassy said in October that the next-generation Trainium 2 chip was ramping up. "We're seeing significant interest in these chips, and we've gone back to our manufacturing partners multiple times to produce much more than we'd originally planned," Jassy said.
Still, Anthropic's Amodei sounded this week like he was hedging his bets.
"We believe that our mission is best served by being an independent company," he said. "If you look at our position in the market and what we've been able to do, the independent partnerships we have Google, with Amazon, with others, I think this is very viable."
Anthropic has raised an additional $4 billion from Amazon, and has agreed to train its flagship generative AI models primarily on Amazon Web Services (AWS), Amazonβs cloud computing division. The OpenAI rival also said itβs working with Annapurna Labs, AWSβ chipmaking division, to develop future generations of Trainium accelerators, AWSβ custom-built chips for training AI [β¦]
Β© 2024 TechCrunch. All rights reserved. For personal use only.
