In today’s edition, we have a scoop on the under-the-radar startup Turing, whose work helped launch ͏ ͏ ͏ ͏ ͏ ͏ |
| Reed Albergotti |
|
Hi, and welcome back to Semafor Tech.
Every day, I read another article about how an AI company scraped some type of data from the web and used it to train a model. It’s true that this is a legal gray area that copyright law hasn’t yet sorted out.
But the moral outrage over this kind of data collection is, in my opinion, misplaced. (And I say this as someone whose book and countless articles have been scraped by AI.)
All that information was needed to help develop current AI models, but it looks like it will be a lot less important in the next phase of this technology.
For the most part, these massive datasets were used in pre-training AI models. That phase of AI development is kind of like the period in which toddlers pick up language and a basic understanding of the world around them by simply being around adults. Exactly what those adults say and do isn’t all that important. They just need to say something and do something.
In other words, my book and all those articles contributed a tiny fraction of a percent to the development of AI toddlers.
We’re well past that phase. Getting AI models to behave in useful ways, like answering questions instead of just completing sentences, only happens after those massive datasets have been processed.
And now we’re heading into a new period of AI development where researchers want more specialized data. Companies are paying PhDs and other experts to produce very specific information for their latest AI models. The hope is that these programs will learn how to reason and make decisions, not just regurgitate raw data.
Nobody knows when the next AI breakthrough will happen. For now, though, it looks like all those petabytes of internet data are less important than the much smaller, more targeted and clearly legal forms of data that may take AI models to the next level.
You can read more about this below and how a $1 billion startup has completely reoriented its business around this phase of AI development.
➚ MOVE FAST: Say hello. Samsung has passed a key hurdle in proving its chips are up to snuff to make Nvidia’s AI processors, Reuters reported. That opens the door for the two firms to sign a supply deal amid high demand for Nvidia’s products. Samsung, the world’s largest chipmaker, needs that win for its manufacturing race against fellow South Korean firm SK Hynix, the leading maker of high-bandwidth memory chips. ➘ BREAK THINGS: Wave goodbye. OpenAI co-founders Greg Brockman and John Schulman have left as the company continues to see turnover at the top. Brockman said he’s taking a long sabbatical and will be back, while Schulman is joining rival Anthropic. Those kinds of changes are common at young companies. But OpenAI’s board fight and allegations of prioritizing profits over safety have led some to speculate that there’s more to these departures than typical startup drama. Steve Jennings/Getty Images for TechCrunch/Wikimedia Commons |
|
European cities are finding creative ways to make use of the heat generated by data centers powering AI. The latest facility built in Paris suburb Seine-Saint-Denis, for example, is keeping France’s Olympics swimming pools warm. Hot air produced in the data center, named PA10, is piped away and used as a source of heating elsewhere to prevent its servers from getting too warm. It was built by Equinix, a US digital infrastructure company, and will reportedly be able to produce as much heat to keep 1,600 homes warm, according to Nicolas Buono, senior director of operations, at Equinix. Pool via ReutersIt’s a cheaper way to keep data centers cool, since it doesn’t require fancy cooling systems that need more energy. But critics believe this deflects attention from soaring energy demands spurred by constructing more data centers for AI, Wired reported. – Katyanna Quach |
|
Turing’s key role in the launch of ChatGPT |
THE SCOOP In early 2022, Jonathan Siddharth, CEO of staffing firm Turing, drove up from Palo Alto for a meeting at OpenAI’s San Francisco office. Siddharth thought he was there to pitch OpenAI on Turing’s specialty: finding high-quality software engineers to work as contractors. Instead, he found himself in a room full of AI researchers who wanted something entirely different: data. OpenAI had found that adding bits of computer code into the datasets used to train its large language model, then GPT-3, helped increase the model’s reasoning ability. That’s why its AI researchers wanted as much computer code as they could get, and they wanted it to be good. They asked Siddharth if he could raise an army of computer coders who would complete specific software engineering tasks so their work could be ingested into OpenAI’s next project: GPT-4. “What I remember distinctly is the scale of their ambition in terms of how big they were thinking,” Siddharth told Semafor in an interview. “Their demand from us was like a crazy high spike in terms of how much data they wanted in how little time.” Courtesy of Turing Turing’s work played an important role in helping OpenAI make a massive leap in performance that blew away the competition and shocked the world when it was released in November 2022 in the form of ChatGPT, according to a former OpenAI employee with knowledge of the matter. It also changed Turing’s business model. In the nearly three years since it began working with OpenAI, Turing has launched a suite of AI consulting services, and counts as clients nearly every major foundation model provider, as well as large companies that want to train their own AI models. The $1 billion startup has also moved beyond coding and is providing specialized data from a wide array of industries. “It’s very impressive, the transition they pulled off,” said Quora CEO Adam D’Angelo, who is an investor in Turing and a board member at OpenAI (he wasn’t involved or aware of the first pitch meeting with Turing). “It’s almost like the first business just got them these capabilities internally around managing large numbers of contractors and engineers. And then that was just the perfect way to build that up to then be able to address this new market, which I think is ultimately going to be much bigger.” How data supplied by subject-matter experts can help usher in the “agentic” era of AI. → |
|
|
“I don’t believe there’s a price in the world that Microsoft could offer us. They offered to give us Bing for free. They could give us the whole company.” |
— Eddy Cue, Apple’s senior vice president of services, on how the iPhone maker still wouldn’t want to use Microsoft’s search engine on its Safari web browser. Cue made these remarks in testimony during the US antitrust trial against Google, which Judge Amit Mehta included in court documents that ruled Google has illegally maintained a monopoly in search. |
|
Courtesy of Sweetspot Government contracting startup Sweetspot, which uses AI to help companies navigate the US procurement process, raised $2.2 million in its initial seed funding round, the company told Semafor exclusively. The round, led by 1984 Ventures, marks a milestone for the five-person startup that came out of Y Combinator in 2023. It hopes to make the sometimes byzantine, $759 billion world of government contracts accessible to anyone. It’s also an example of the new generation of AI-enabled startups that can get by with fewer employees and less funding. One of Sweetspot’s five employees is an intern. The company launched about a year ago with a search engine that scours federal contract databases and has since expanded to mine state and local projects. The company’s customers include AI startup Groq, defense tech company Vannevar Labs, security firm Strider Labs and Flexport, according to Sweetspot co-founder Sachin Subramanian. He told Semafor the idea is to essentially create a “TurboTax” for government contracts, making it easy for everyone from general contractors to web developers to find and fulfill government contracts. Sweetspot is in a race with another Y Combinator startup, GovDash, which was founded in 2022 and is more heavily funded. |
|
Pool via Reuters Elon Musk’s antics have always raised eyebrows, but he’s been particularly busy as of late. He filed not one but two lawsuits this week. The first one is against OpenAI, reviving a legal action he had dropped in June. Musk sued the startup and its co-founders, CEO Sam Altman and President Greg Brockman, accusing them of manipulating him into helping launch OpenAI as a non-profit. Musk said he feels “betrayed” since it’s now a for-profit business (overseen by a non-profit) and believes the change is a breach of contract. The second case is against a group of advertisers for refusing to pitch their goods and services on X. Despite Musk’s fierce defense of free speech, he claims in his lawsuit that they are unfairly boycotting his social media platform and likely violating antitrust laws. Meanwhile, five US state officials urged him to fix misinformation being spread by Grok, X’s AI chatbot, that could unfairly influence the upcoming presidential election. In a letter, the secretaries of state from Minnesota, Pennsylvania, Michigan, Washington, and New Mexico highlighted that Grok had inaccurately claimed the ballot deadline for voting in a handful of states had already passed. He has also landed himself in hot water with the British Prime Minister Keir Starmer after Musk declared that “civil war is inevitable” in response to the recent anti-immigration riots in the UK. The country’s parliament also wants to summon him for X’s role in spreading misinformation after three children were killed in a knife attack that sparked the violence. On top of that, the British Computer Society and The Chartered Institute for IT have called for the country’s broadcasting regulator, Ofcom, to fine X for stoking fears. If that wasn’t enough, he’s currently in the middle of a custody battle with his ex-girlfriend and musician, Grimes, over their three children and was called to court in Austin last week, Business Insider reported. — Katyanna Quach |
|