The week in AI: Decoding DeepSeek

In partnership with

Welcome to The Dispatch! We are the newsletter that keeps you informed about AI. Each Thursday, we aggregate the major developments in artificial intelligence - we pass along the news, useful resources, tools and services; we highlight the top research in the field as well as exciting developments in open source. Even if you aren’t a machine learning engineer, we’ll keep you in touch with the most important developments in AI.

NEWS & OPINION

When hype, however justified, meets reality: everything DeepSeek

-------------------------

What a messy whirlwind of a week for the AI community around the landmark release of DeepSeek’s R1 model. Last Thursday, we had a breakdown of the model and called it “probably the most important (or at least shocking) development in open source AI to date” and that the precise implications were “hard to say at this point, but potentially profound.” We didn’t need to wait too long for validation - on Monday, DeepSeek sent US stocks into a tailspin with Nvidia alone losing $600B in market value (the highest single day loss, by more than double, in history).

Commenting the same day, President Trump noted that it was a wake-up call for the US tech industry, but that in the spirit of competition and increasing the efficiency of AI overall, it “… could be very much a positive development.” The dust has not settled yet by any means, but one week later we can cut through some noise and make a more nuanced analysis.

Starting with the model itself:

• Let’s not mince words: R1 is state-of-the-art competitive. This model release is probably the first time we’ve seen a widespread “benchmark by use case” scenario across digital publications, technical blogs, and social media. That’s exciting, because the established benchmarks most of these companies rely on to flex their muscle leave much to be desired. By virtually any metric, R1 is a frontier LLM (while still suffering from the same errors, hallucinations and blips as other models at a similar rate).

• Its democratization has really charged up the general public - think Robin Hood comes to the AI space. The fact that R1 came on the heels of OpenAI’s $200/mo subscription for their more advanced models was not lost on anyone. As we noted last week: it’s fully open source, the API is dirt cheap, the largest model in the family can run locally on a relatively modest cluster and the distilled models can run on consumer-grade hardware (and even your phone). The distilled models aren’t great overall, but pretty good at reasoning, math and coding - especially for their size.

• R1 showcased a novel training paradigm, notably its successful implementation of pure Reinforcement Learning for initial reasoning capability development with R1-Zero - bypassing traditional supervised fine-tuning. This is coupled with a sophisticated multi-stage pipeline, including rule-based rewards and bootstrapped SFT data. Why is this a big deal? In simpler terms, they found and made public (more on that later) a smarter and much cheaper way to teach AI to reason, likely changing how we build these models going forward. If you were one of the skeptics who thought LLMs could only ape their training data, think again.

Moving on to DeepSeek, and how reality delivered a swift slap to the face:

• DeepSeek founder Liang Wenfeng made a bet that a small team of young, hungry, Chinese-educated AI researchers and engineers uninhibited by traditional business structures could innovate best - and he was right. This team is cracked. It’s also important to understand that this approach puts DeepSeek far outside the traditional Chinese innovation system. They’re an outlier, not a representation.

• You might be asking yourself why DeepSeek chose to open source their work. While the company claims it’s about an open source culture and the fact that they don’t need more money, it’s more likely they are attempting to commoditize their complement, a well-known strategy successfully employed many times in open source’s history.

• Unfortunately for them, it turns out you need more than just a small team of envelope-pushing engineers to sustain a successful AI product at scale. DeepSeek’s virality made it #1 on app stores in both the US and China, the first time any Chinese app has ever done so. And then some cracks started showing.

  • They had to limit registrations due to large-scale cyberattacks. For days now, their API has been inaccessible, although most recently they state that it should be fixed soon.

  • Wiz research exposed a massive DeepSeek database, which allowed full control over database operations - the leak was over a million lines of log streams and contained chat histories, secret keys, backend details and other sensitive information.

  • Ireland and Italy already sent data watchdog requests to DeepSeek. Italy blocked them in app stores and has filed a complaint related to how they handle data relating to GDPR laws. Knowing the EU, more countries will likely follow suit.

  • David Sacks, Trump’s AI and crypto “czar,” and OpenAI have confirmed there’s evidence that DeepSeek used OpenAI’s AI models to train R1, a process that violates OpenAI’s terms of service and ‘equates to theft’. There’s a level of irony to that allegation that verges on artificial superirony - and today Microsoft (OpenAI’s biggest investor) added to insult by hosting R1 on Azure in spite of this infraction. Ouch. Even so, DeepSeek will likely faced increased barriers to their training methodology from competitors going forward.

  • DeepSeek faces national security scrutiny from all sides. The Chinese Communist Party has had their thumb on Chinese Big Tech for years; DeepSeek has operated under their radar but that will likely not continue, moving forward. The US National Security Council is reportedly reviewing the security implications of using DeepSeek’s models, and the US Navy has already banned the use of DeepSeek even for personal use.

The stock market, Stargate, US tech firm responses, export controls on AI chips, and everything else:

• The stock market plunge caused by DeepSeek was surely in part due to the recent $500B Stargate Project for AI infrastructure in the US - why invest to that scale, when DeepSeek just proved you can do so much more with so much less? This knee-jerk reaction misses a crucial point: the future of AI compute is very, very broad. While DeepSeek's model is a landmark, consider the data intensity of emerging AI applications. Robotics, with multi-sensory input and real-time processing needs, advanced simulations and video analysis, next-generation video games (the gaming industry alone is larger than all of film, television and music combined) aiming for unprecedented realism, just to name a few – these are areas poised to supercharge compute demands. The market's focus on LLM efficiency as a negative signal for compute demand makes no sense.

• Microsoft CEO Satya Nadella further pointed to Jevons Paradox - the idea that technological progress that increases the efficiency of resource use can lead to an increase in the overall consumption of that resource, rather than a decrease. In 19th century Britain, steam engines were becoming significantly more efficient at burning coal to produce power. The intuitive expectation was that with more efficient engines, Britain would use less coal. However, Jevons observed the opposite: coal consumption actually increased rapidly.

• While there were some early reports of panic among US tech giants, it appears that they are now doubling down on hardware as a moat. It might be their last one; but it might be a good one. The reality is that no one yet knows the true effect of export controls the US has levied on China. The semiconductor industry boasts the most complex supply chain in the world - and despite the announcements of export controls having begun back in 2022, Dutch company ASML supplies virtually every advanced chipmaker in the world with the equipment required to make cutting edge AI chips; their compliance with the US restrictions has gone into full effect only a few weeks ago.

• To emphasize how hard this is to scrutinize, on one side you have reputable American think-tanks claiming that DeepSeek highlights that export controls are not only not working, but possibly counter-productive. Meanwhile, if you get a bit more granular with the issue you can find DeepSeek’s CEO himself noting in a translated interview (read: Part 3) that the company’s barrier has never been about money or talent - rather, the US export controls are the problem for the company. Who do you think you should believe?

• Anthropic CEO Dario Amodei provided more context in a posted essay. According to Amodei, these controls aren’t about preventing China from obtaining a few advanced chips (and DeepSeek’s chip acquisitions were already known or suspected), but rather to ensure they can’t secure millions of chips needed to reach AI parity with the US in 2026-2027. Amodei believes that DeepSeek's success is also a reminder of how effective and ambitious China’s AI talent is, and that the export controls are, right now, the only major countermeasure to that strength.

• Additionally, Amodei notes that: A) DeepSeek did not pioneer the training methodology for R1 - using reinforcement learning to train models to generate chains of thought became a new focus of scaling for frontier AI companies in 2024 starting with OpenAI’s o1. This scaling mission is still in the infancy stages, and Amodei argues whoever takes advantage quickest is probably going to reap major rewards. B) The media might be overstating American inefficiency: “DeepSeek does not "do for $6M what cost US AI companies billions". I can only speak for Anthropic, but Claude 3.5 Sonnet … cost a few $10M's to train (I won't give an exact number).”

There’s probably more on the DeepSeek story that we missed, and the talking points are endless. In the wake of all this, how might Trump feel about Meta releasing Llama 4 into the wild? At what point does open source itself become a potential national security issue, and what would that even mean? Will o3 make R1 look pedestrian? Is now the “last best time” to invest in Nvidia after such a steep drop, or have they been destined for a bubble pop all along?

Impressions of OpenAI’s Operator are coming in - an agent that can go to the web to perform tasks for you

-------------------------

On any other week, OpenAI’s first entry into the agentic space would have been the top story with a full breakdown. Since we don’t have the space to do so this week without getting clipped and flagged by Gmail and other inboxes, we’ll keep it brief and point you to some resources:

• Since HubSpot is sponsoring us this week, it seems appropriate to first highlight CTO Dharmesh Shah’s post on his personal newsletter - which happens to be focused on AI agents - about how Operator works and the broader implications for the industry. Like many, he is cautiously optimistic.

• But Operator is only a research preview, and has a long way to go. Hunter.io founder Antoine Finkelstein outlines some of Operator’s limitations and weaknesses right now. Still, even he closes on an optimistic note.

• OpenAI put out an official introduction to Operator and Agents on YouTube.

• Also on YouTube, creator Matthew Berman made a great compilation of use cases and reactions coming through social media. The AI Advantage also just put out a great video on “10 actually useful things” you can do with it today.

• Box CEO Aaron Levie claims that Operator can “basically straight up use cloud software to do anything” while showcasing it working on his platform.

Again, Operator is still in research phase - unless you’re willing to pay the $200/mo to test with it under tempered expectations, you’re probably safe just keeping tabs on it for now. Still, there has been much doubt circulating about agentic capabilities; this looks like a good entry for OpenAI and they will continue to maintain their first-mover’s advantage with the data they reap from it. Operator will improve quickly.

MORE IN AI THIS WEEK

• OpenAI launches ChatGPT Gov for US government agencies

• Which AI to use now: an updated opinionated guide

• Microsoft makes OpenAI’s o1 reasoning model free for all Copilot users

• US Copyright Office releases a new report establishing firm guidelines on AI-generated works (Matt Wolfe breaks it down on YouTube; there’s a few muddy areas around ‘partial’ copyrights)

• Cleveland police used AI that is explicitly inadmissible in court to justify a search warrant - and it has derailed a murder case

• Paul McCartney warns UK’s proposed AI law change could rip off artists

• Macron-backed French AI “Lucie” launch is a total disaster

• I want my AI to get mad

Need a personal assistant? We do too, that’s why we use AI.

Ready to embrace a new era of task delegation?

HubSpot’s highly anticipated AI Task Delegation Playbook is your key to supercharging your productivity and saving precious time.

Learn how to integrate AI into your own processes, allowing you to optimize your time and resources, while maximizing your output with ease.

Get the free guide here.

TECHNICAL NEWS, DEVELOPMENT, RESEARCH & OPEN SOURCE

• Alibaba releases Qwen2.5-Max: exploring the intelligence of large-scale MoE models - and Qwen2.5-VL, the new flagship vision-language/early agentic model that can interact with computers and phones

• Twitter founder Jack Dorsey’s fintech company launches codename goose: an open-source AI agent framework for devs to build and deploy AI assistants

• DeepClaude: harness the power of DeepSeek R1's reasoning and Claude's creativity and code generation capabilities with a unified API and chat interface

• Hong Kong Univ. of Science and Tech unveil YuE: an open-source AI music “lyric-to-song” generation model

• Hugging Face announces SmolVLM 256M and 500M: the world’s smallest vision language models that maintain competitive performance; also launches four new serverless inference providers for Hub, enabling direct model deployment and faster inference

• DeepSeek staying busy: releases Janus Pro 7B, a multimodal image generator (fully open source, not quite state of the art but quite good)

• PyTorch releases an update with FP16 support on X86 CPUs, improving AI inference and training performance (previously GPU-only)

That’s all for this week! We’ll see you next Thursday.