GitHub goes multi-model, a move with large implications

What happened

On 29 October, Microsoft's division GitHub, which provides tools and assistance for software developers, announced that its Copilot assistant (1.8mn users), designed to help developers improve their coding efficiency, would start integrating Anthropic's Claude 3.5 Sonnet and Alphabet's Gemini 1.5 Pro Large Language Models (LLMs). The company previously relied on a single provider, OpenAI, of which the integration of additional models has also been announced to supplement the current offering. The change will be effective in the coming weeks.

Impact on our Investment Case

What this means for OpenAI

With the launch of ChatGPT, OpenAI instantly acquired a dominant status rarely seen in any other field, exemplified by ChatGPT becoming the de facto synonym of Artificial Intelligence. This allowed the company to raise capital at increasingly high valuations ($80bn in February 2024, $157bn in October 2024), and gave its CEO Sam Altman free reign to transform what was initially a non-profit organization into a would-be Big Tech. However, cracks have been appearing of late: Apple pulled out of the latest funding round at the last moment; after some serious drama leading to Altman being fired and then hired again, several prominent figures (CTO, Chief Scientist, AGI adviser) have left the organization; and the partnership with Microsoft is showing signs of weakness.

GitHub opening to new providers may be seen as another sign of relative decline, especially if this shift is extended to other Microsoft products. It either means that OpenAI is no longer considered the gold standard, or that Microsoft has enough doubts to want to start hedging its bet. In any case, a clear window of opportunity is opening for competition.

What this means for the sector

In our view, at the risk of repeating ourselves, it means that end applications have definitely taken the driver's seat. LLMs like GPT or Claude are ultimately tools coming in different flavors. Users do not use them for the sake of using them, but because they want to accomplish a precise objective. In the process, given the complexity of LLMs and the tasks they can tackle, it is normal to see a model outperforming in certain operations but underperforming in others. The next logical step would be to select the model on demand depending on the workload, which is precisely what GitHub now allows the user to do.

We are, therefore, increasingly approaching the moment where LLMs will become commoditized, or at least "bland"-enough to be relegated to the background. The actual star of the show will be the applications making the best of them through thorough implementation. There will still be money to be made in LLMs, but this will heavily rely on the access to data allowing them to be heavily specialized. Due to the rapidly growing capabilities of real-time search (e.g., through RAG), generalist models may ultimately become commonplace enough to lead to massive consolidation, or even to be open-source (see Meta's approach with its Llama model).

What this means for Atonra

We do not only talk about AI and invest in corresponding stocks. We are actively working on implementing AI technologies at the core of our workflow and investment process to optimize them as much as possible. In this regard, our current implementation approach is perfectly in phase with GitHub's: we have indeed developed an open architecture allowing to integrate as many LLMs as wanted, and to facilitate switching from one to the other depending on the workload. By doing so, we can select the best tools currently available (our money goes right now to OpenAI and Anthropic) while not closing the door to further innovation on the supply side, or evolution on our side.

Our Takeaway

This move by GitHub is both a testament of the frantic sector innovation leading to a relative decline of OpenAI, and a confirmation that the sector is entering a new cycle past the initial technology trigger. It confirms our view that value ultimately resides in end-applications, and that players higher in the value chain need a strong moat if they want to thrive. Our investment process is precisely designed to identify such opportunities, and has led us to shift the portfolio towards applications for some quarters already. We will keep going this way, identifying the most promising segments and the players capable of capturing the growth where it actually manifests itself.