AI Roundup 103: The DeepSeek edition
January 31, 2025.
By now, you’ve undoubtedly heard about DeepSeek, the Chinese AI company that has caused massive shockwaves in the tech world. As much as I'd like to cover something else, the nonstop parade of DeepSeek stories continues, so let's look at some second-order events in this week’s news roundup.
DeepSeek: Integration
As DeepSeek's chatbot has suffered from outages, other companies have been busy integrating and replicating the company's models: from Azure, to Perplexity, to Hugging Face.
Why it matters:
Because DeepSeek has published both its models and its research, other teams have been able to build and host their own versions of the model.
The releases seem to vindicate Meta's strategy of giving away its AI models for free, even though some reports say the tech giant is now scrambling internally.
Interestingly, Microsoft CEO Satya Nadella seems to have been ready and waiting for something like DeepSeek - the company was able to add R1 to their Azure platform in only 10 days.
Elsewhere in the FAANG free-for-all:
Google is launching Ask for Me, an AI that calls businesses about hours and pricing, and Gemini in Sheets for Google Workspace users.
Microsoft's AI business revenue has reached a $13B annual run rate, up 175% year-over-year.
And Mark Zuckerberg committed to investing hundreds of billions in AI infrastructure, starting with up to $65 billion in 2025 alone.
DeepSeek: Regulation
Although many see DeepSeek as a threat to US AI companies, it also serves as powerful leverage in their attempts to influence the regulatory environment.
The big picture:
The heads of both OpenAI and Anthropic have now cited DeepSeek as a reason to expand domestic investment and curtail foreign chip access.
Even though DeepSeek's breakthroughs have been enormously impressive, the company is bottlenecked by AI chips, suggesting that future advances will be tricky without new sources of compute.
President Trump has called this week's media frenzy a "wake-up call," vowing to impose tariffs on foreign chips in the near future.
Elsewhere in AI DeepSeek anxiety:
DeepSeek left one of its critical databases exposed, leaking over 1M records, including system logs, user prompts, and API keys.
According to cybersecurity firms, "Hundreds" of companies and government agencies globally are moving to block DeepSeek due to data leak concerns.
DeepSeek's privacy policy states that user-provided content is stored on servers in China - a requirement for Chinese technology companies.
And one study shows that DeepSeek's chatbot had an 83% fail rate on news-related prompt and repeated false claims 30% of the time.
DeepSeek: Distillation
Microsoft and OpenAI are investigating whether a group linked to DeepSeek improperly accessed and used OpenAI's technology to help develop their rival R1 model.
Between the lines:
OpenAI claims to have evidence of "distillation" - a technique that uses outputs from more advanced models to train smaller ones - being used by DeepSeek in violation of OpenAI's terms of service.
Distillation represents a growing challenge of protecting AI intellectual property, as smaller players leverage outputs from established models to avoid expensive training costs.
Of course, there’s some irony in the situation: as 404 Media puts it, "OpenAI is furious DeepSeek might have stolen all the data OpenAI stole from us."
Elsewhere in OpenAI:
OpenAI is in early talks to raise $40B at a massive $340B valuation, as revenue from the $200/month ChatGPT Pro plan reportedly surpasses Enterprise sales.
The company announced that US National Laboratories will use its o1 models for research and work with Microsoft to deploy them on supercomputers.
A new ChatGPT Gov launched, intended for US government use. More than 90,000 government employees have already generated millions of prompts.
And OpenAI faces challenges transitioning to a for-profit structure, particularly regarding how to value Microsoft's stake.
Elsewhere in frontier models:
Mistral AI launched Small 3, a 24B-parameter model optimized for latency.
Alibaba claims its new Qwen 2.5-Max model outperforms GPT-4 and DeepSeek-V3 across most benchmarks.
The Qwen team introduced Qwen2.5-VL, a model family capable of controlling PCs and phones.
And DeepSeek also launched Janus-Pro, an open-source text-to-image model (like DALL-E and Stable Diffusion).
Things happen
HudZah, a 20-year-old, builds a nuclear fusor with Claude's help. DeepSeek may complicate hopes for venture-backed IPO rebound. Microsoft makes OpenAI's o1 model free for all Copilot users. AI is creating a generation of illiterate programmers. Mistral AI loses ground to US rivals and DeepSeek. Block introduces on-device, open-source AI agent called Goose. AI models now good enough for PhD-level analysis. The bear case for Nvidia. LLM-written posts steadily increasing on Physics Forums. Ex-OpenAI researcher terrified by AI development pace. US Copyright Office: AI-assisted content can be copyrighted. Countries use AI tools to bolster cyberattacks. DeepSeek's chatbot fails 83% of the time on news-related prompts. Character AI claims First Amendment protection in lawsuit over user's suicide. I do not want AI to "polish" me. Bill Gates talks social media, crypto, and AI in memoir interview. When AI promises speed but delivers debugging hell. Using AI for coding: A journey with Cline and LLMs. Hundreds of companies block DeepSeek due to data leak concerns. AI slop, suspicion, and writing back. California AG tells AI companies everything might be illegal. Using AI to develop a fuller model of the human brain. Knowing less about AI makes people more open to using it. AI tarpits to trap and trick AI scrapers. I trusted an LLM, now I'm on day 4 of an afternoon project. Vatican note on the relationship between AI and human intelligence. US could encourage future DeepSeeks by funding the NAIRR. Bridgewater founder Ray Dalio says AI has fueled a stock market bubble.
Is this the singularity? I know it's a trope and a joke, but seriously: I can't keep up with all this without using AI tools to put them into context. Close, right?