Housekeeping
Welcome to the last roundup of 2024! I’m planning a couple of year-in-review style posts for next week, including potentially a year-long roundup, but this is the final “regular” roundup of the year. Happy holidays, and see you in 2025!
Flash Thinking
This week, Google joined the AI reasoning race with its Gemini Flash Thinking Experimental model (and continued its parade of model launches from last week).
Why it matters:
Flash Thinking is part of a broader industry shift toward "reasoning" AI models, such as OpenAI's o1 and Qwen's QwQ, that can self-check their work and potentially provide more reliable results.
Some, including former OpenAI cofounder Ilya Sutskever, see these models as the future of AI, as a lack of high-quality training data forces the industry to consider new inference paradigms.
Yet despite the promise of reasoning models, early testing reveals limitations (like miscount letters in words), and their high costs raise questions about their long-term viability.
Elsewhere in Google:
Veo 2, a video generation model capable of creating clips over two minutes long at resolutions up to 4K.
Whisk, an image generator that uses other images as prompts and utilizes the latest Imagen 3 version.
Agentspace, a Google Cloud feature enabling enterprises to create and deploy AI agents for various business functions (including NotebookLM Plus).
Interactive Audio Overviews in NotebookLM, allowing users to converse with AI "hosts" of podcast-like overviews.
Support for third-party tools to Gemini Code Assist, allowing access to real-time data from external apps.
And the FACTS Grounding benchmark for evaluating the factuality of LLMs, with Gemini 2.0 Flash Experimental topping the leaderboard.
Elsewhere in the FAANG free-for-all:
Meta teases an AI editing tool for Instagram powered by Movie Gen AI, allowing users to change outfits, backgrounds, and more in videos using text prompts.
Apple is reportedly in early-stage talks with Tencent and ByteDance about integrating their AI models into iPhones sold in China, as ChatGPT remains unavailable in the country.
And Meta rolls out new Ray-Ban Meta features, including "live AI" for videos, live translations for Early Access Program members, and Shazam integration for US and Canadian users.
House of Cards
Meanwhile, despite its many new launches, OpenAI is still fighting political and legal battles on multiple fronts that could determine its future as a company.
What to watch:
On the one hand, OpenAI is continuing its push to become a for-profit, even as Meta urges California's Attorney General to stop the transition.
The for-profit shift is also the target of an ongoing lawsuit from Elon Musk, as new court documents argue that Musk wanted to run a for-profit version of OpenAI himself.
On the other hand, the company (along with the rest of the tech industry) is cozying up to Trump ahead of his inauguration, including by making a seven-figure donation to the inaugural fund.
OpenAI, in particular, is at risk with the new Trump administration, as its former co-founder Musk continues to wield power as the co-head of the DOGE.
Elsewhere in OpenAI:
Projects, a new ChatGPT feature for organizing files and chats into folders, available to Plus, Pro, and Team users with plans to extend to free users soon.
1-800-CHATGPT, a phone line offering 15 minutes of free monthly access via phone call in the US or WhatsApp messaging globally.
ChatGPT Search is now available to all users, not just paid subscribers, and can be set as the default search engine.
And o1 is now available via API for select developers, along with new versions of GPT-4o and GPT-4o mini as part of its Realtime API.
Elsewhere in AI anxiety:
Anthropic demonstrated "alignment faking" in Claude 3 Opus to show how developers could be misled about an LLM's true alignment.
The BBC complained to Apple after Apple Intelligence's notification summary falsely suggested the BBC had published an article about Luigi Mangione shooting himself.
And universities are facing an "AI cheating crisis" as students use AI tools to complete assignments and exams.
Copyright conflict
The UK government is proposing new rules requiring AI companies to be more transparent about their training data, while offering limited copyright exemptions that creative industries can opt out of.
Between the lines:
Under these rules, tech companies would need to disclose what content they use for training and clearly label AI-generated outputs - a significant shift toward transparency.
But creatives (including Paul McCartney and Kate Bush) have already resisted the proposal, saying it forces creators to "opt out of shoplifting" and views it as a fundamental threat to copyright protection.
Other companies are testing out alternate licensing strategies: YouTube creators can choose which AI companies can train on their videos, while Shutterstock is introducing a "research license" model that’s cheaper than a commercial license.
Elsewhere in AI geopolitics:
The US' proposed AI chip framework would block adversaries entirely and give others quotas based on their US alignment, threatening Nvidia's expansion.
Outgoing US DHS secretary Alejandro Mayorkas says Europe's "adversarial" relationship with tech companies is hampering a global approach to AI regulation.
Malaysia, which has secured billions in AI and cloud infrastructure investments, launches a national AI office for policy and regulation.
The UK's AI Safety Institute delays plans to open a US outpost as it faces Trump's "starkly" different regulatory approach.
China plans to establish standards for LLMs and AI risk assessment by setting up a new committee (with representatives from Baidu and Peking University).
Things happen
Bengaluru's growth is straining its infrastructure. Ornithologists can use BirdVoxDetect to track migratory songbirds. Truth Terminal, an AI bot on X, sparks debate and boosts memecoins. US policies could deprive companies of AI talent from China. Salesforce to hire 2,000 for AI agent software sales. Non-AI startups struggle to raise Series B amid AI boom. Companies are using AI agents for various tasks. Donald Trump's AI czar urges tech startups to boost US competitiveness. Q&A with Fei-Fei Li on visual intelligence and her startup. Musk's DOGE role conflicts with OpenAI regulation push. AI app contributes to suicide reduction in Mexico's Yucatán state. App Store saturation makes finding good AI apps harder. Ex-ByteDance intern wins NeurIPS Best Paper after dismissal. OpenAI whistleblower found dead in San Francisco. Klarna stopped hiring a year ago due to AI investments. Claude becomes chatbot of choice for AI insiders. Sundar Pichai on Google's AI strategy. Anthropic's Clio identifies new threats and generates insights. Microsoft bought 485K Nvidia GPUs in 2024. xAI rolls out upgraded Grok-2 to all X users. GitHub offers free Copilot with limited features. 90% of AI training datasets come from Europe and North America. AI-driven electricity demand could strain US and Canadian grids. CAA partners with YouTube to remove AI-generated content. Workers adopt generative AI faster than companies can issue guidelines. Ad giants merge to prepare for AI-driven upheaval. AI will be dead in five years. AI is lying about its power. You're sleeping on AI SWEs. AI companies may shift to smaller, specialized models.