AI Roundup 099: Apocalypse Not
January 3, 2025.
Apocalypse Not
With many 2024 retrospectives coming out (including one from yours truly), one of the most surprising things about the year was how little we heard about AI doomers.
Zooming in:
2023 had a ton of AI doom momentum: calls for AI moratoriums, AI safety summits, and leading figures discussing their p(doom).
Now, in 2025, there's none of that - the biggest battle against AI regulation was SB 1047, and the technologists won (for now). Even the "AI election misinformation" boogeyman was fairly underwhelming.
Of course, just because AI isn't going to kill us all doesn't mean it doesn't have downsides right here, right now. And I fully expect more AI anxiety in the coming year.
Speaking of which:
Anthropic agreed with publishers to stop Claude from providing lyrics to their songs or generating new lyrics based on copyrighted content, pending a court ruling.
Experts warn of a rise in personalized phishing emails targeting corporate executives with personal details likely obtained via AI analysis of online profiles.
A tipster was arrested after federal authorities discovered AI-generated child exploitation images and plans to create virtual reality CSAM.
Waymo robotaxis, now ubiquitous in parts of California, seemingly fail to stop for pedestrians using crosswalks unless they are far into the road.
And Meta expects AI-generated characters to generate and share content, and eventually exist on their platforms similarly to regular accounts.
DeepSeek
Two weeks ago, Chinese AI company DeepSeek released V3 of their open chat model - and early results are astounding the AI community in terms of performance and cost.
Why it matters:
The latest model release beats GPT-4o/Claude 3.5 Sonnet on multiple benchmarks while being 1) open weights, 2) trained with 10x fewer resources, and 3) an order of magnitude cheaper.
DeepSeek previously released R1, a reasoning model that has mostly flown under the radar despite previously outperforming OpenAI's o1 on some benchmarks (of course: benchmarks aren't everything).
There does appear to be evidence, though, that ChatGPT-generated answers have contaminated DeepSeek's training data (either accidentally or intentionally).
Elsewhere in 中國:
Ant Group's payment app Alipay added an AI-powered image search feature as it competes with rival Tencent's WeChat to become a super app.
Alibaba cut prices on its visual language model Qwen-VL by up to 85%, following earlier price cuts of up to 97% on Qwen models amid growing AI competition in China.
And Chinese AI chip designer Cambricon's stock jumped 383% in 2024, reaching a $37B market cap, driven by the push for AI chip localization in the country.
Public benefit
OpenAI is moving forward with its plans to become a for-profit organization, starting with a letter from its board explaining the rationale.
Between the lines:
It appears that the company will split into a non-profit and for-profit; the non-profit will receive shares in the for-profit, though it will only be in charge of charitable initiatives.
Many are still skeptical of the transformation, and it will have to survive the courts - Elon Musk received new support for his lawsuit last week from third-party filings.
There are also new details about OpenAI's deal with Microsoft - reportedly, OpenAI defined "achieving AGI" as the point at which its AI systems generate $100B in profits.
Things happen
CFTC v. Gemini trial delayed to January 21. Nvidia invested $1B across 50 funding rounds in 2024. Paris vies with London as Europe's top AI hotspot. The biggest AI flops of 2024. AI platforms unlikely to become major traffic source for media outlets. Microsoft working on adding more AI models to 365 Copilot. AI hallucinations supercharge the scientific method. A look at more challenging AI evaluations emerging. AI companies dominate forum traffic. Spotify is full of AI music. Does current AI represent a dead end? Google's pressure-filled 2024 under Sundar Pichai. Building AI Products: Back-End Architecture. A look at IDF's Unit 8200 and AI-derived intel. Deepfake mentions on X correlated with new image generation models. Defense industry poised for deal surge in AI, drones, and space systems. Nvidia completes $700M acquisition of Run:ai.
