Wall Street tempers its AI expectations, antitrust enforcers tally a win against Google, and NTIA weighs in on open weights

Sign up
Worth Knowing

Wall Street Recalibrates Its AI Expectations: Some investors are starting to temper their initially bullish outlook on artificial intelligence. The recalibration, which sent ripples through the stock market in recent weeks, has been driven by an apparent mismatch between the costs and benefits of AI systems. Tech giants like Microsoft, Alphabet, and Amazon are pouring massive amounts into AI startups, infrastructure, and R&D, but some industry analysts are beginning to question whether these substantial investments will yield commensurate rewards. “Tech giants and beyond are set to spend over [$1 trillion] on AI capex in coming years, with so far little to show for it … will this large spend ever pay off?,” Goldman Sachs wrote in a recent report. Some prominent economists like Daron Acemoglu have long questioned whether AI systems will deliver the broad-based productivity gains promised by the technology’s boosters, and the economic data from the last two years has done little to refute their claims. Even prominent labs like OpenAI are struggling to turn a profit amid soaring labor, infrastructure, and energy costs.

While generative AI systems have demonstrated payoffs in narrow areas such as code development and customer service, analysts say, companies have generally been slow to adopt the tech, and ongoing challenges like algorithmic bias and hallucinations have cast doubt on how well AI tools can scale. The growing skepticism about AI has contributed to a general downturn in tech stocks. Between July 10 and August 5, Google, Microsoft, Amazon, saw their share prices plunge more than 10%, and chipmakers like Nvidia and Intel lost more than a quarter of their market value.

Amid this volatility, large tech companies have sought to allay investors’ concerns and carry on as planned with their long-term AI ambitions. In quarterly earnings calls, Google announced it would pour more than $24 billion into AI infrastructure in the second half of 2024, and both Microsoft and Meta alerted investors to expect similarly substantial investments in the months ahead.
The AI Models Keep on Rolling: Despite Wall Street’s recent bearishness, the technology’s leading developers have continued to roll out a steady stream of powerful new models.
  • On July 18, OpenAI released GPT-4o mini, a lightweight multimodal model that is significantly cheaper to run than its more powerful counterpart, GPT-4o. OpenAI said the model excels in mathematical reasoning and coding, and comes equipped with safety measures that protect against potential misuse. The release came less than a week after The Washington Post reported that OpenAI illegally restricted employees from voicing concerns about the potential safety risks of the company’s products to regulators. The alleged practice came to light after whistleblowers filed a complaint with the Securities and Exchange Commission.
  • On July 23, Meta released its largest and most powerful language model to date. Billed as “the first frontier-level open source AI model,” Llama 3.1 reportedly performed around the same level as OpenAI’s GPT-4 and Anthropic’s Claude 3.5 on a variety of benchmarks. Performance aside, Meta also claims the model can be run at about half the cost as OpenAI’s GPT-4o.
  • Not to be outdone, Google DeepMind released an experimental version of its powerful Gemini 1.5 Pro model which some in the AI community have heralded as the most powerful model on the market. The model, available through the Gemini API and Google AI studio, has surged to the top of the LMSYS leaderboards, a community-run ranking of model performance. Google DeepMind also released a pair of mathematical reasoning models—AlphaProof and AlphaGeometry2—that performed at the same level as the silver medalist in the 2024 International Mathematical Olympiad. Google DeepMind leveraged a combination of the Gemini large language model and their AlphaZero algorithm to achieve this feat.

  • More: AI companies promised to self-regulate one year ago. What’s changed? | Who Will Control the Future of AI? | Exclusive: Anthropic wants to pay hackers to find model flaws
Grok Spreads Election Misinformation: Government officials are up in arms after the large language model run by X (the platform formerly known as Twitter) spent more than a week spreading misinformation about the 2024 election. Shortly after President Joe Biden dropped out of the presidential race on July 21, the company’s model, Grok, started sharing false information that Vice President Kamala Harris would be ineligible to appear on the ballot in nine states. The error was reportedly not corrected until July 31. The incident prompted several secretaries of state to pen an open letter demanding X CEO Elon Musk make changes to “ensure voters have accurate information in this critical election year,” and specifically to have Grok direct users asking election-related questions to official resources. Researchers including CSET’s Josh Goldstein have long raised concerns that AI systems could be used to orchestrate cheaper, more expansive misinformation campaigns. Beyond the Grok fiasco, election officials around the globe find themselves battling against deepfakes and other misleading content generated by AI.
Government Updates

Antitrust Enforcers Log A Win Against Big Tech: On August 5, a federal judge determined that Google had illegally abused its market power to shut out rivals from the market for search engines. “Google is a monopolist, and it has acted as one to maintain its monopoly,” Judge Amit Mehta wrote in a 277-page ruling. The landmark decision concluded the first antitrust case against a major technology company in more than 20 years. The case hinged largely on the vast sums that Google paid companies like Apple and Samsung to make its search engine the default on their smartphones and web browsers. According to Judge Mehta, the agreements offered Google “access to scale that its rivals cannot match” and left other search engines “at a persistent competitive disadvantage.” Legal scholars say the decision could make judges more receptive to the government’s arguments in other ongoing antitrust cases against tech giants like Apple, Amazon, and Meta. The remedy for Google will be decided in a separate case scheduled for September. Some experts have questioned whether any remedy could change the market for internet search all that much, and others have suggested generative AI firms like OpenAI actually present a bigger risk to Google’s search business than federal regulators.

The ruling could open the door to future antitrust enforcement in the AI industry. In a joint statement released on July 23, competition authorities in the U.S. and Europe committed to proactively policing the AI industry and addressing risks to competition “before they become entrenched or irreversible harms.” Already, the U.S. Federal Trade Commission has opened a probe into partnerships between major AI labs and cloud providers, and the U.S. Department of Justice is investigating Nvidia for potentially abusing its dominance in the chips market. Prior CSET research has explored the ways breaking up tech companies like Google could impact the Defense Department’s access to AI systems.

NTIA Weighs in on Open Weights: On July 30, the National Telecommunications and Information Administration (NTIA) published a report recommending the Biden administration support open models as a way to democratize the AI ecosystem. “‘Open-weight’ models allow developers to build upon and adapt previous work, broadening AI tools’ availability to small companies, researchers, nonprofits, and individuals,” agency officials wrote. The report marks an important development in the debate over the potential costs and benefits of open AI models. Proponents of open weights emphasize the ways that transparency could galvanize innovation and promote a more community-oriented approach to AI safety, but critics warn that open models allow adversaries to more easily adopt AI capabilities and potentially use the technology for nefarious purposes. While the report comes out in favor of open models, it also acknowledges the risks  such technologies pose and urges policymakers to monitor the AI ecosystem and implement restrictions when appropriate. Among other things, the report recommends that the government research the safety and downstream uses of powerful open models and develop risk-related benchmarks to inform possible future policy changes. The report cited work from CSET researchers Thomas Woodside, Helen Toner, and Kyle Miller.

The NTIA report stands in stark contrast to other recent AI governance measures, which seek to tighten regulators’ control over the development and distribution of powerful models. For example, California’s SB 1047, which currently awaits a final vote in the state assembly, would hold developers liable for certain downstream uses of their models if those models exceed a certain size, a rule that would likely limit the distribution of certain open models. Many AI developers have spoken out against the California bill, claiming its provisions would stifle innovation and burden developers with costly compliance measures.  

Job Openings

We’re hiring! Please apply or share the roles below with candidates in your network:
  • Survey Researcher: We are currently seeking candidates for a Survey Researcher position. In this role, you will lead survey design and execution in collaboration with CSET analysts (Senior Fellows, Research Fellows, Research Analysts) across CSET’s lines of research. You will sit on CSET’s data team and will work closely with data scientists, software engineers, a translation lead, and data research analysts to produce, collect, and analyze original survey data. This is a hybrid role that combines a mix of research methods and data analysis skills. Strong applicants will have experience with survey design, online survey distribution platforms (e.g., Qualtrics), as well as project management, and qualitative and quantitative data analysis. Apply by September 9, 2024
Please bookmark our careers page to stay up to date on all active job postings. You can also subscribe to receive job announcements by updating your subscription preferences in the footer of this email.

In Translation
CSET’s translations of significant foreign language documents on AI

Disruptive Innovation: Interim Measures of Shanghai Municipality for the Administration of Disruptive Technology Innovation Projects. This document details how Shanghai Municipality oversees state-funded “disruptive technology” projects. The regulation provides detailed instructions on how to deal with projects that are behind schedule, failing, or fraudulent from the get-go, suggesting that these are common problems for Chinese government-financed technology projects.

Xi Touts Chinese Science: Xi Jinping: Speech at the Nationwide S&T Conference, National Science and Technology Awards Conference, and the Conference of Academicians of CAS and CAE (June 24, 2024). This document includes the text of Chinese President Xi Jinping’s speech at a major Chinese S&T conference in June 2024. In his remarks, Xi recaps many of China’s scientific triumphs of 2023 but also warns that the nation still has weaknesses in areas such as original innovation and talent.

If you have a foreign-language document related to security and emerging technologies that you’d like translated into English, CSET may be able to help! Click here for details.

What’s New at CSET

REPORTS

PUBLICATIONS

EMERGING TECHNOLOGY OBSERVATORY

UPCOMING EVENT

  • Join CSET on August 29 for Building the Tech Coalition, a half-day conference featuring distinguished speakers discussing the military’s application of AI and software. This event is sponsored by Domino Data Lab.

IN THE NEWS

What We’re Reading

Report: The Root Causes of Failure for Artificial Intelligence Projects and How They Can Succeed, James Ryseff, Brandon De Bruhl, and Sydne J. Newberry, RAND (August 2024) 

Article: Unpacking New NIST Guidance on Artificial Intelligence, Gabby Miller, Tech Policy Press (August 2024)

Article: The AI Safety Debate Is All Wrong, Daron Acemoglu, Project Syndicate (August 2024)