AI Geekly: Markets are Clueless

Welcome back to the AI Geekly, by Brodie Woods, brought to you by usurper.ai. We may have been off for the last two weeks, but that doesn’t mean we haven’t been following the developments!

Our Alignment Check note from two weeks ago couldn’t have been more timely! Readers will recall we discussed the strange bedfellows of the Open Source Community + Chinese tech corps + Meta created by the embrace of each of these groups of open research concepts (but for different reasons as we discussed). Equity markets have suddenly clued in to the now expensive realization that OpenAI, Anthropic, and the predominantly Closed Source Big U.S. tech companies are not the only game in town. Paired with concerns about higher efficiency in training, some wonder whether the big data center build-outs are even needed if players like DeepSeek-AI/High Flyer Capital can build o1-level models for a song. We’ll dig into this more in this week’s Geekly at the end of the note. If you want to read that immediately, just press here. But if you have a moment, there’s some more AI news you might find relevant, particularly OpenAI’s new Operator Agent that can actually complete tasks.

♫Operator, Could You Help Me Place This Call?♫
OpenAI releases its new agent

OpenAI finally took the wraps off its agent offering, dubbed Operator (big Jim Croce fans, we’re sure!) The new AI tool (current only available to ChatGPT Pro users at $200/mth) allows users to enter text prompts to have the Operator AI perform tasks for the user. We’ve been experimenting with real world use cases and have been fairly impressed. While not as versatile as Anthropic’s Computer Use Agent, which can use a full computer UI, Operator can only use a browser (like Google’s Mariner). To put it to the test, we had it research and summarize information about the four recently released AI models we wanted to mention. Watch the video above to see how it did in practice and see below for the output. We think you will agree that it is quite impressive (remember that this is the worst it will be, it only gets better from here).

Here is what Operator compiled about the models in this test on a single attempt:

DeepSeek Janus multimodal model

• Creators: Developed by DeepSeek-AI (High Flyer Capital)

• Novelty: Janus-Pro is an advanced version of Janus with optimized training, expanded data, and increased size, enhancing multimodal understanding and text-to-image stability.

• Comparison: It significantly improves text-to-image instruction following compared to previous models.

MiniMax

• Creators: Developed by a team including Minimax Aorian Li, Bangwei Gong, Bo Yang, and others.

• Novelty: MiniMax-01 series offers superior long-context processing using lightning attention and Mixture of Experts (MoE), scaling efficiently to 456 billion parameters.

• Comparison: It provides superior capabilities compared to top-tier models in handling longer contexts.

DeepSeek R1 reasoning model versions

• Creators: Developed by DeepSeek-AI.

• Novelty: DeepSeek-R1-Zero and DeepSeek-R1 are trained via large-scale reinforcement learning, demonstrating remarkable reasoning capabilities without supervised fine-tuning.

• Comparison: DeepSeek-R1 achieves performance comparable to OpenAI-01-1217 on reasoning tasks.

DeepSeek v3

• Creators: Developed by DeepSeek-AI.

• Novelty: DeepSeek-V3 is a Mixture-of-Experts (MoE) model with 671B parameters, using Multi-head Latent Attention and auxiliary-loss-free strategies for efficient inference and training.

• Comparison: It outperforms other open-source models and rivals leading closed-source models in various benchmarks.

DeepSeek’s Turntable Stop
Markets miss the mark

There’s lots to unpack about the DeepSeek story. For a quick recap, DeepSeek-AI was founded in 2023 and is backed by High-Flyer Capital. A proponent of Open Source AI, the China-based AI lab has developed several models in its DeepSeek text and Janus image model series. Its open-source models compete toe-to-toe with the best closed models from Open AI, Anthropic and other AI leaders. DeepSeek caught the attention of investors yesterday (Jan 27), partly due to its app reaching #1 on the iOS app store, driving a major decline in equity markets erasing over $1 Trillion in value. We’ll unpack each of the key themes relevant.

Valuation and Speculation: We DO NOT provide investment advice. Arguably, some of this decline is justified, as names like Broadcom, Oracle, and Micron may have been overbought on marginal AI exposure relative to their overall business. Nvidia let $593 Bn out of the balloon, but having half a trillion in market cap to lose in the first place is certainly a champagne problem... There is no question that investing in AI names has become speculative, one need only look at the trading multiples of nearly every large AI company to see lofty valuations many standard deviations above the norm.

Training Efficiency: One of the concerns that pummeled AI hardware and energy companies was learning that DeepSeek R1 and v3 were both trained on relatively little hardware (2,048 Nvidia H800s, a sanctions-weakened version of the H100) and for relatively little money ~$5.6 mm. The theory goes that if High Flyer was able to train o1 and 4o level models for 1/100th of the cost and in a fraction of the time previously believed to be standard, then demand for Nvidia’s GPUs could dry up. The reality is that EVERYONE is trying to train more efficiently, OpenAI included. DeepSeek employed novel techniques to efficiently use hardware due to necessity, as do many in the Open Source community with limited access to hardware. We’re hesitant to call High Flyer / DeepSeek-AI GPU Poor given their balance sheet but they are making do with limited resources in a way that benefits other GPU Poor players in the Open community.

Compute Matters: As training efficiency improves through innovations in approach, architecture, and hardware, compute demand will likely remain extremely high. This is largely due to the ongoing development of Test Time Compute, the process of letting an AI model use compute to iterate and reason through its response to provide a more accurate output is extremely compute intensive. We expect slack created by more efficient training to get picked up both by a higher volume of players training and fine-tuning models, and increased demand for inference compute for deployed models.

What this is: Meta’s Chief AI Scientist, Yann LeCun sums it up best in his LinkedIn post. This isn’t China surpassing the US in AI. For one thing, High-Flyer is a modestly sized quant fund, not a state-run AI lab. The Open Source community creates a fertile environment for collaboration and innovation, and it is attracting the top talent. The market for ideas is highly competitive and right now it looks like Open Source is beginning to inch ahead.

About the Author: Brodie Woods

As CEO of usurper.ai and with over 18 years of capital markets experience as a publishing equities analyst, an investment banker, a CTO, and an AI Strategist leading North American banks and boutiques, I bring a unique perspective to the AI Geekly. This viewpoint is informed by participation in two decades of capital market cycles from the front lines; publication of in-depth research for institutional audiences based on proprietary financial models; execution of hundreds of M&A and financing transactions; leadership roles in planning, implementing, and maintaining of the tech stack for a broker dealer; and, most recently, heading the AI strategy for the Capital Markets division of the eighth-largest commercial bank in North America.