AI Roundup 091: Search wars

November 1, 2024.

Nov 01, 2024

∙ Paid

Search wars

Perplexity has quickly developed a reputation as the leader in the "AI search engine" space. But other AI giants are now setting their sights on Perplexity's turf.

Between the lines:

Yesterday, Google announced Search Grounding for the Gemini API, which generates AI responses grounded in search results, and OpenAI released ChatGPT Search, which incorporates real-world data and links.
Even Meta appears to be getting in on the search engine action, though it appears to mostly be in response to its dependence on Google and Microsoft.
However, there are still many open questions about what it means to shift our current search engine paradigm, not least of which is: "How are people going to make money without ads?"

The SearchGPT Paradigm

Charlie Guo

Aug 1

Read full story

Elsewhere in the FAANG free-for-all:

Microsoft's AI business is on track to surpass a $10B annual revenue run rate in Q2, its fastest product category to reach this milestone.
Microsoft disclosed $13B in funding commitments to OpenAI, with losses from its stake contributing to a $683M expense on Q1 investments.
Microsoft has delayed the rollout of its controversial Recall feature to December to "refine the experience".
Google's AI Overview summaries will roll out in over 100 countries and territories, reaching over 1 billion monthly users.
And Google plans to preview an AI tool codenamed Jarvis, which can control a user's browser to complete tasks like booking flights, as early as December 2024.

Careless Whisper

For many AI chatbots, call recorders, and note-taking apps, OpenAI's Whisper is the de facto model for transcribing audio. But new AP reporting suggests that the model suffers from serious hallucinations.

Why it matters:

Whisper and similar models have already experienced massive adoption because they're faster and significantly cheaper than human-based alternatives.
But like other generative AI models, there's still the chance that the model hallucinates. While most errors are small, occasionally, entire sentences are inserted that don't exist.
The issue is worsened by the fact that hospitals and medical centers are already using speech-to-text models to transcribe doctor's visits, meaning potentially dire consequences for hallucinated symptoms or treatments.

Elsewhere in AI anxiety:

Anthropic urges governments to take urgent action on AI policy in the next 18 months, suggesting principles for AI regulation that support innovation.
Chinese and Israeli academics predict the AI boom will increase global e-waste by 3% to 12% by 2030, potentially adding up to 2.5 million metric tons annually.
And a UK judge has sentenced a 27-year-old man to 18 years in prison for using AI tool Daz 3D to create child sexual abuse imagery, marking a landmark prosecution over deepfakes.

Elsewhere in OpenAI:

In a Reddit AMA, Sam Altman says they won't be releasing anything specifically called "GPT-5" this year.
OpenAI has been working with Broadcom on its first in-house AI chip for inference, built by TSMC.
OpenAI's CFO reveals ChatGPT has 250M weekly active users and 75% of revenue comes from consumer subscriptions.
And OpenAI introduces SimpleQA, a new dataset of 4,326 questions to evaluate LLM performance on factual answers.

Meta has a lot going on

I'm continually struck by how prescient (in hindsight) Mea's decisions to invest in GPUs and a strong AI team were (ten years ago, it wasn't clear why a social network needed either). In a more detailed breakdown, Ben Thompson lays out the case for why Meta is very well positioned in the short, medium, and long term.

How it works (in theory):

Short term: Meta can capture some of the massive VC dollars through its ad platform as companies try to prove the ROI exists for generative AI hype.
Medium term: Meta can use AI to generate perfectly personalized content for users' feeds while transforming every piece of content (both AI-generated and authentic) into potential advertising inventory through AI-powered object recognition and linking.
Long term: Meta is building projects like Orion (AR glasses), where AI-powered generative UI will enable frictionless interactions while also solving VR content creation challenges through AI-generated environments - all of which can be monetized through its advertising infrastructure.

Elsewhere in Meta:

Meta is training Llama 4 models on a cluster of 100K+ H100 chips, which Zuckerberg claims is larger than competitors' setups.
Mark Zuckerberg says Meta is "executing quite well" in building AI infrastructure, as the company raises its 2024 capex guidance.
Meta is collaborating with Wonik and GelSight to develop tactile sensors for AI, including a robotic hand and tactile fingertips.
The company disagrees with OSI's definition of open-source AI, as Llama doesn't fit due to commercial use restrictions.
Meta AI has surpassed 500M monthly active users in less than a year, while Threads approaches 275M MAUs.
And the company has struck a deal with Reuters to use its news content for real-time answers in Meta AI chatbot.