AI Roundup 110: Siri stumbles

March 21, 2025.

Mar 21, 2025

Siri stumbles

Last year, Apple countered critics who saw it as unprepared for the generative AI era by announcing Apple Intelligence. Now, that same criticism is back as the company delays key Siri upgrades.

The big picture:

Several long-time technology bloggers have called out Apple for delaying key AI features and pulling an ad that prominently showcased those features.
The growing narrative is that Apple's AI strategy has deep issues, including buggy features, missed deadlines, and "ugly" delays.
In response, the company is shaking up its leadership this week: Vision Pro creator Mike Rockwell will take over Siri development from John Giannandrea.

Elsewhere in the FAANG free-for-all:

Google is expanding its healthcare initiatives with new Search and Android features, including medical records APIs and plans for TxGemma AI models for drug discovery.
Google enhances Gemini with a new Canvas feature for creating and sharing writing and coding projects, along with Audio Overview capabilities.
Meta's Llama models have reached a milestone of 1 billion downloads since their 2023 launch.
And users have discovered that Gemini 2.0 Flash is particularly effective at removing watermarks from images, unlike some competing models.

openai.fm

OpenAI released three new voice and transcription models (offshoots of GPT-4o), and an interactive demo to try them out at openai.fm.

What's new:

Two of the models are for for transcription, with improvements to language recognition and accuracy compared to the original Whisper models. Unlike Whisper, these models won't be open-sourced.
The new speech model, on the other hand, now has an "instructions" parameter, meaning you can direct the model with specific emotions, tone, or accents.
According to OpenAI, the models were also trained using reinforcement learning and distillation techniques - meaning the research advances from GPT-4o and o1 are enabling better "narrow" models, starting with audio.

Elsewhere in model mayhem:

OpenAI also released o1-pro, a more powerful version of o1, in the API - priced at $150/1M input tokens.
Aardvark, an AI weather prediction system that uses thousands of times less computing power than current methods.
Stability AI's Stable Virtual Camera, a non-commercial AI model that creates 3D videos from up to 32 2D images.
Mistral's new 24B-parameter model outperforms Gemma 3 and GPT-4o-mini while running on just 32GB RAM.
Roblox open sources Cube 3D, its first AI foundation model for generating 3D objects.
Baidu releases Ernie X1 for reasoning tasks and upgrades to Ernie 4.5, both available for free.
And Anthropic adds web search to Claude 3.7 Sonnet in preview for US paid users.

Peer pressure

AI companies are stirring new controversy by submitting AI-generated papers to a major AI conference (ICLR), with some failing to disclose the synthetic nature of their submissions to peer reviewers.

Between the lines:

While one company, Sakana, disclosed its AI-generated papers, others like Intology and Autoscience submitted papers without disclosure.
There's a race to build the first "AI scientist" - but it's worth being skeptical about. Sakana's AI only had one of three papers accepted, even as it made "embarrassing" citation errors.
The incident highlights growing tensions between AI companies using peer review as a benchmarking tool and academics who view it as an exploitation of volunteer labor - peer reviewers are mostly unpaid.

Elsewhere in AI anxiety:

A US appeals court three-judge panel ruled that AI-generated works cannot be copyrighted, since the Copyright Act requires human authorship.
Ben Stiller and over 400 Hollywood creatives sent an open letter urging President Trump to prevent tech companies from exploiting copyrighted works with AI.
And Republican Jim Jordan has subpoenaed major tech companies including Apple, Amazon, Meta and OpenAI over allegations of Biden administration pressure to use AI for censorship.

GTC 2025

Nvidia held its GTC 2025 conference this week, with several announcements, partnerships, and products.

Everything Nvidia:

Nvidia's GTC 2025 announcements focus on addressing the scaling challenges of AI pre-training, post-training, and inference working together.
The company updates its lineup with DGX Station reservations for the DGX Spark box and new RTX PRO workstation GPUs.
Nvidia introduces AI-Q Blueprint, a framework for connecting knowledge bases to agents, alongside new Llama Nemotron AI models.
The company unveils two new AI chips: Blackwell Ultra, which will ship in H2 2025, and Vera Rubin, which will have a custom CPU, in H2 2026.
Nvidia's new Dynamo operating system claims to boost token output by 30x per GPU when running DeepSeek-R1.
GM partners with Nvidia to implement Drive AGX and Omniverse technologies across driver assistance systems and factory operations.
Nvidia acquires Gretel, a synthetic data tools provider, for more than its previous $320M valuation.
And Bloomberg looks at the four companies, including Nvidia and TSMC, that dominate the AI supply chain with a combined market cap over $4T.