TheChek News Flow

Of all of OpenAI’s rivals, Anthropic seems to be worrying the artificial intelligence leader the most, especially its progress in key areas like automatic code writing.

The Takeaway

• OpenAI execs have expressed concerns about Anthropic’s business growth

• Anthropic’s annualized revenue from AI coding has grown 10 times in last three months

• Anthropic has gotten more comfortable with releasing experimental products

OpenAI has a lot of rivals nipping at its heels. None of them has given the company’s executives the jitters quite like Anthropic has, though.

Earlier this fall, OpenAI leaders got a shock when they saw the performance of Anthropic’s artificial intelligence model for automating computer programming tasks, which had gained an edge on OpenAI’s models, according to its own internal benchmarks. AI for coding is one of OpenAI’s strong suits and one of the main reasons why millions of people subscribe to its chatbot, ChatGPT.

OpenAI leaders were already on edge after Cursor, a startup OpenAI funded last year, in July made Anthropic’s Claude model the default for Cursor’s AI coding assistant instead of OpenAI’s models, as it had previously done, according to an OpenAI employee. In a podcast in October, Cursor co-founder Aman Sanger called the latest version of Anthropic’s model, Claude 3.5 Sonnet, the “net best” for coding in part because of its superior understanding of what customers ask it to do.

For Anthropic, its success in coding seems to be having a business impact. The annualized revenue the company generates from customers who use its models for software development and code generation has increased by 10 times over the last three months, Kate Jensen, Anthropic’s head of growth and revenue, said in an interview.

To counter that momentum, OpenAI has scrambled to improve the coding capabilities of OpenAI’s own models, said a person who was involved in the response. In recent weeks, OpenAI leaders including Chief Product Officer Kevin Weil have pointed to Cursor, among other examples, as reasons they’re concerned about the growth of Anthropic’s business, another employee said.

OpenAI’s concerns about Anthropic show how rapidly the conversational AI business is changing. While OpenAI remains the dominant force among the current generation of AI startups, its rivals have gradually eaten into its technological lead, including by poaching some of its employees. In Anthropic’s case, its business has grown faster than its leaders had projected, helping it cut into OpenAI’s dominant share of the conversational AI business.

An OpenAI spokesperson reiterated the strength of its consumer and enterprise businesses.

Anthropic is hardly the only competitor OpenAI needs to keep an eye on. There’s Google and Elon Musk’s xAI, which has turned heads in the AI business recently with how quickly it built a giant data center in Memphis. And while Microsoft and Amazon have partnered with and made large investments in OpenAI and Anthropic, respectively, they also compete with them by building rival AI services.

To be sure, OpenAI, which is five years older than Anthropic, still maintains a comfortable lead in revenue over Anthropic. It has been on pace to generate roughly $4 billion in revenue this year, at least five times more than the revenue Anthropic previously projected for the year. And OpenAI dwarfs it in other ways, raising a total of $20 billion at a peak valuation of $157 billion, to Anthropic’s $11 billion in funding and peak valuation of $18 billion.

OpenAI is also in a much stronger financial position than its rival because it shares a smaller percentage of the revenue it generates with its cloud provider, Microsoft, than Anthropic shares with its main cloud partner, Amazon. Technology from OpenAI and Anthropic powers key products sold by Microsoft and Amazon.

OpenAI is attempting to shift its governance structure to that of a for-profit business outside the control of its nonprofit board. One potential option under discussion is a for-profit benefit corporation, which rivals such as Anthropic and xAI are using.

OpenAI and Anthropic are also incinerating cash on the order of billions of dollars each this year due to the high cost of developing and running their technologies. OpenAI has expensive ambitions to develop its own data center chips and possibly other hardware so it can reduce its dependence on outside suppliers, which will likely keep it hunting for capital well into the future.

Still, Anthropic has one self-imposed guardrail that has prevented it from moving even faster in its AI development in the past: its focus on safety. That term refers to the efforts at AI companies, including at OpenAI, to ensure their technologies don’t make serious mistakes or take actions that jeopardize human lives—in extreme cases, through the development of biological weapons or by initiating nuclear strikes.

All seven of Anthropic’s co-founders came from OpenAI, but they left it in late 2020 over concerns about AI safety. One of those co-founders, Anthropic CEO Dario Amodei, has said the startup developed an AI chatbot in the summer of 2022 but decided against publicly releasing it, instead continuing safety testing.

Months later, OpenAI released ChatGPT, creating a huge stir among consumers and the tech industry. Anthropic launched its Claude chatbot four months later.

Recently, Anthropic has gotten bolder about tweaking its bigger competitor. In October, after a flurry of exits from OpenAI, including Chief Technology Officer Mira Murati, advertisements for Anthropic’s Claude AI began appearing in the San Francisco International Airport with a not-so-subtle jab: “The one without all the drama.”

Anthropic also seems to have gotten more comfortable with releasing experimental products. In October, it launched a product that allows its conversational AI to take over customers’ computers to automate tasks such as building a website or editing a spreadsheet. It did so despite a risk of cyberattacks that could exploit the technology, as the company said in a blog post.

The move got an amused reception inside OpenAI. At a recent meeting, OpenAI leaders skewered Anthropic’s decision to launch the product, despite the risks and Anthropic’s high-minded rhetoric about AI safety, according to an employee.

Someone Else’s Vision

The bad blood between the founders of Anthropic and OpenAI started long before they split over safety issues.

As OpenAI’s vice president of research, Amodei led the development of OpenAI’s GPT-2 and GPT-3 models and also co-authored a seminal AI paper on reinforcement learning from human feedback, alongside other researchers from OpenAI and Google DeepMind. This technique allows humans to give feedback on model responses and has driven much of the improvement in conversational AI.

At OpenAI, Dario and Daniela Amodei—his sister, who is now Anthropic’s president—clashed with other executives, especially CEO Sam Altman and President Greg Brockman, over who should lead certain projects, along with safety issues, according to former employees.

In early 2019, Brockman was working on AI that could play the online battle game Dota 2. Later, Brockman attempted to join the OpenAI research effort, overseen partly by Dario Amodei, to develop language-generating GPT models, which would eventually become the foundation for blockbuster products like ChatGPT, two former employees said.

However, the Amodei siblings prevented Brockman from joining the project, in part because he had a reputation for being difficult to work with and making last-minute changes to software code without telling others, they told other employees. Brockman later got involved in the GPT project after the Amodeis and other OpenAI researchers left the company to start Anthropic.

In the months leading up to their departure, the rift between the Amodeis and others at OpenAI became so deep that Dario Amodei created private Slack channels for researchers that Altman, Brockman and others weren’t invited to, one of the former employees said.

In an apparent reference to his time at OpenAI, Dario Amodei on a November podcast said differing opinions on how to build AI safely led to the schism.

“If you have a vision for how to do it, you should go off and you should do that vision,” he said. “It is incredibly unproductive to try and argue with someone else’s vision.”

Simple Features

At Anthropic, there are subtle differences between the company’s priorities and those of OpenAI, which could partly explain why Anthropic has made inroads with business customers.

Because its Claude chatbot gets only a fraction of the usage ChatGPT does, Anthropic has focused its research efforts more on helping it better sell its Claude large language models to businesses and developers, employees say.

‘The one without all the drama.’

One result of that is that Anthropic has been slower than OpenAI to release multimodal models—which can ingest and produce information in nontext formats, such as images, videos and audio—and reasoning models that can solve more complex, multistep problems, an ability useful for science researchers.

In the eyes of Anthropic leaders, those tools go far beyond the simple applications of the technology most current AI customers are focused on today, such as summarization, content generation and customer service bots, according to a person who has spoken to them. The relatively low usage of reasoning models among ChatGPT customers appears to support that idea.

Instead, Anthropic is devoting more resources to launching features that business customers are clamoring for, such as larger context windows, which have the ability to upload large amounts of information that AI models can use to answer questions. It is also focused on making sure its models can use external tools such as databases and application programming interfaces that help companies build apps incorporating real-time, proprietary information, said the person who has spoken to Anthropic executives.

Anthropic’s Jensen said most companies are building AI into products that have been around for a while. That requires features like larger context windows, which may not sound very sexy but can have a big impact.

These are “things that feel simple but change the game,” Jensen said.

Model Battle

Anthropic’s emphasis on practical improvements for businesses has helped it win over some prominent customers, including cloud-based collaboration startup Airtable.

Howie Liu, CEO of Airtable, said Claude’s larger context windows allow his employees to upload lengthy transcripts from sales calls so the model can better highlight details of the conversation such as its tone and client complaints.

At times, Claude’s responses can feel more humanlike than those of other LLMs, Liu added. Airtable has also tested OpenAI’s models in the past and continues to use them for other applications, Liu said.

Legal research firm LexisNexis, meanwhile, uses Claude for approximately 60% of its AI features, said Chief Technology Officer Jeff Reihl. Many of these are related to drafting or analyzing legal documents, where large context windows are helpful, Reihl said. The remainder of the features use OpenAI models like GPT-4o and open-source models such as Mistral AI, he added.

In October, Intercom announced that Fin, its AI chatbot for resolving customer support tickets, would be powered by Anthropic’s Claude models rather than OpenAI’s LLMs. The customer support startup said in a blog post that Claude allows its agents to answer “more questions, more accurately, with more depth and more speed,” helping it resolve 51% of customer support tickets without any additional tweaks to the model.

That’s up from the 23% resolution rate when Intercom launched Fin powered by OpenAI models in March 2023, an improvement that is likely due to other factors in addition to the change in models. (Anthropic has also announced it will use Fin as its customer service AI agent.)

Code Advantage

One of Anthropic’s biggest edges over OpenAI has been in automating coding tasks.

Before this summer, Augment, a coding assistant startup that has raised more than $250 million in funding, only used free, open-source models for its products, said Scott Dietzen, the company’s CEO.

After Anthropic released Claude 3.5 Sonnet in June, Augment started paying for it because they believed it was a “breakthrough model,” superior to other models it tested, including those from OpenAI, Dietzen said.

“There is still a material gap currently between open source with our post-training and what Sonnet is capable of for use cases that deserve more inference time [and] deeper understanding,” Dietzen said.

Similarly, the customers of Sourcegraph, another coding assistant that relies on Claude 3.5 Sonnet as its default model, choose to stick with the Anthropic model two-thirds of the time rather than switch to those from OpenAI, Google or others, said its co-founder and CEO, Quinn Slack.

In October, even OpenAI’s most prominent ally, Microsoft, began allowing developers to use non-OpenAI models in its GitHub Copilot coding assistant, including from Anthropic.

Altman, for his part, hasn’t said much publicly about Anthropic or the Amodeis, but he noted last month at an event that his rival “has a model that is great at coding for sure, and it’s impressive work.”

‘All the Drama’

There are recent signs that Anthropic has decided to go on offense against OpenAI.

The company has significantly beefed up its sales team: There are now more than five times as many people working on its enterprise sales team as there were a year ago, said Jensen. The pitch those salespeople are using has changed too.

As of early this year, Anthropic salespeople had a pretty standard message for potential customers: They typically acknowledged that most developers were using models from OpenAI and others, and they simply argued that Anthropic’s models would be cheaper or better for specific purposes such as building custom models, said a person who heard the pitch.

By June—after the company released its latest Claude model—it began taking a more aggressive approach in those sales pitches, this person said. Anthropic’s sales reps began telling potential clients its models had overtaken OpenAI’s multiple times this year with its new releases and bragged about the average scores of Claude models on several popular AI benchmarks, the person said.

Not everybody in the AI business agrees that benchmarks accurately reflect how models perform on realistic, day-to-day tasks. But a number of customers of AI models have praised Anthropic’s technical progress this year.

At a Goldman Sachs event in November, Aravind Srinivas, CEO of AI search engine startup Perplexity, called Claude 3.5 Sonnet a “real breakthrough, not just [for] our product but [for] a lot of other products in the market” after it came out in the summer.

The model, Srinivas said, was “much, much better at reasoning.”

Echo AI, whose generative artificial intelligence tools allow companies to analyze customer conversations, has been bought by Calabrio, a maker of software for contact centers that’s owned by private equity firm Thoma Bravo, according to Echo CEO Alexander Kvamme.

Seven-year-old Echo was one of 78 startups included on The Information’s list of generative AI startups that were likely acquisition targets because they had not raised money in at least two years or because they were developing cost-intensive products like foundation models or robots.

Echo, with 40 employees, had raised $35 million from investors including Kleiner Perkins, Insight Partners and Quiet Capital, most recently in April 2021. Echo’s employees started integrating their technology into Calabrio’s offerings this week, said Kvamme on LinkedIn. Terms of the buyout weren’t disclosed.

Gapping down
In reaction to earnings/guidance:

Other news:

Analyst comments:

Gapping up
In reaction to earnings/guidance:

Other news:

Analyst comments:

From grand complications to digital classics, here are HTSI’s top timepieces of 2024

Western democracies drive a global rise in resource nationalism

Businesses worldwide face increased risk from a global rise in protectionism as countries scramble to secure access to the minerals critical to battery manufacturing and the energy transition, according to new research. 

Rising geopolitical tensions have fuelled a rise in state intervention and protectionism “not seen since the first half of the 20th century in western democracies”, global risk intelligence company Verisk Maplecroft said on Thursday. 

The change has been particularly acute in Europe and North America, with governments in both regions seeking to secure their access to critical minerals such as lithium and copper — the supply chains for which are dominated by China — the researchers said.

Increased tensions in the sector have worsened in recent weeks with critical minerals becoming an ever more closely watched geopolitical football. This month, China banned shipments to the US of several crucial minerals and metals in retaliation for new export controls imposed by the Biden administration designed to target Beijing’s development of artificial intelligence. Earlier this year, a coalition of western nations including the US and UK announced financing plans for minerals projects in an effort to diversify away from China.

“The fracturing geopolitical landscape and the fallout from major shocks like the pandemic and Russia’s invasion of Ukraine have spurred an acceleration of policies aimed at acquiring the minerals needed to power the tech and defence industries, as well as the green transition to bolster energy security,” said Jimena Blanco, chief analyst at Verisk Maplecroft. 

“State focus on supply chain security has opened the door for companies to take advantage of attractive incentive schemes, but geopolitical divergence could increasingly limit opportunities to allied or friendly jurisdictions,” she said. 

According to the researchers’ latest resource nationalism index — a quarterly analysis that measures government control of economic activity in the mining and energy sectors — 72 countries out of the 198 assessed had seen a “significant increase” in interventionist and protectionist policies over the past five years.

Venezuela, Russia and Mexico were judged to be the three countries where businesses faced the greatest risks of state intervention and expropriation in the sectors.

But the analysts said that the risk scores for Germany, Spain, the UK and Poland had all worsened significantly since 2019, with Germany registering the largest drop of any country during the period. It has come as a consequence of protectionist moves by Berlin such as the seizure of Russian energy assets following the country’s invasion of Ukraine, and the offering of subsidies to boost domestic mineral processing and manufacturing, they said.

More broadly, the analysts pointed out that European and North American governments had taken steps to shore up their domestic mining and energy industries and restrict foreign investment from rivals with policies including US President Joe Biden’s Chips and Science Act. 

Forty-one countries that were responsible for 41 per cent of global mineral output were now classified as being either “high” or “very high” risk for protectionist policies, the researchers said. That was an increase from 30 countries five years ago.

“The most likely scenario is that western nations will increasingly use a mix of trade and investment policies, along with stricter sustainability standards, to restrict trade with rivals and push for localised supply chains,” said Blanco.

Heightened risks across multiple jurisdictions exacerbated the overall challenges faced by companies and investors, given the complex and cross-border nature of many critical mineral supply chains, the group said. For example, a mineral may be mined in one country but processed in another and sold to a manufacturer operating in a third.

The analysis considered countries’ protectionist and interventionist policies, state participation in resource extraction and instances of direct and indirect expropriation, such as asset nationalisation or regulatory changes that make doing business in the sectors less profitable.