AI Roundup 107: GPT-4.5
February 28, 2025.
GPT-4.5
OpenAI launched GPT-4.5, its largest and most compute-intensive model yet. However, the launch came with some mixed messaging about the new model's capabilities.
Why it matters:
Despite using much more compute than GPT-4, the model underperforms benchmarks compared to recent reasoning models, though produces fewer hallucinations and is generally a better communicator.
But the model is incredibly expensive to run - between 15-30x more than GPT-4o - raising questions about its long-term viability and the future of “bigger is better” AI strategies.
It’s available today for Pro subscribers, with additional rollouts coming soon - Sam Altman has said the company is “out of GPUs.”
Elsewhere in reasoning:
As OpenAI's "last non-chain-of-thought model," GPT-4.5 is a transition point as the company shifts focus to reasoning models and hybrids.
The model release follows Claude 3.7, which we discussed earlier this week, and Grok 3. Both already outperform o1, OpenAI’s publicly released flagship AI model.
Suffice it to say; we’re likely going to see a lot of coding and math progress in a short amount of time, considering how quickly state of the art is advancing.
Elsewhere in frontier models:
Tencent released Hunyuan Turbo S, a new AI model designed for instant responses that claims to match DeepSeek's V3.
IBM unveiled its Granite 3.2 LLM family featuring reasoning capabilities, document vision processing, and time series forecasting.
Hume introduced Octave, a text-to-speech LLM that creates customizable AI voices with emotional expression capabilities.
Alibaba released Wan 2.1, an open-source family of AI models for image and video generation available globally through ModelScope and Hugging Face.
And DeepSeek released not one, but five new open-source repos this week, as part of its Open Infra initiative.
Alexa+
Amazon is finally getting its AI act together with the upcoming launch of Alexa+, a $20/month subscription that brings generative AI capabilities to the smart assistant.
Between the lines:
Amazon has completely redesigned Alexa, aiming for more natural conversations and complex tasks - Alexa+ can autonomously navigate websites and interact with third-party services.
Amazon's approach is “model-agnostic”, combining its own Nova model with third-party models like Anthropic's, plus partnerships with major news outlets and service providers.
The delayed launch (originally targeted for early 2024) highlights the technical challenges of integrating generative AI into existing assistants, with Amazon taking extra time to avoid releasing a subpar product.
Elsewhere in the FAANG free-for-all:
Google's new Gemini Code Assist offers individual users 180K free code completions monthly, vastly outpacing GitHub Copilot's free tier.
Meta is developing a standalone AI app and testing paid subscriptions as Zuckerberg aims to dominate the AI space.
And xAI's new Grok 3 voice mode introduces multiple personalities including controversial "Unhinged" and "Sexy" options.
DeepSeek's great leap forward
In the aftermath of DeepSeek's groundbreaking R1 release, the company has been accelerating not only its own releases but also the Chinese AI ecosystem as a whole.
The big picture:
While it initially planned to release R2 (R1's successor) in May, the company is now reportedly trying to launch it as soon as possible.
While limiting direct profits, DeepSeek's free model distribution strategy is democratizing AI access across China and creating a robust ecosystem that advances Beijing's AI ambitions.
Ultimately, the state's enthusiastic embrace of DeepSeek signals China's determination to consolidate its position in the global AI race through a national champion.
Elsewhere in AI geopolitics:
Microsoft called on Trump to ease US export restrictions on data center chips for allied nations, including India, Switzerland, and Israel.
China is accelerating AI development by pooling resources from both state and private sectors, including startups like Infinigence AI.
Trump directed CFIUS to limit Chinese investments and consider outbound restrictions in strategic sectors like chips and AI.
Apple committed to US investment of $500B+ over four years, including AI server production in Texas and 20,000 new jobs, following Tim Cook's meeting with Trump.
Elsewhere in AI anxiety:
OpenAI discovered evidence of Chinese surveillance operations using AI tools to identify anti-China social media posts in Western countries.
Researchers found that AI models trained on insecure code showed misalignment across unrelated topics.
YC faced backlash and removed social media posts after its backed startup Optifye.AI demonstrated AI monitoring of factory workers.
And a Pew survey reveals 52% of US workers worry about AI's workplace impact, while only 36% feel hopeful about its future use.
Things happen
The debate about whether AI models can truly reason. AI coding tools will likely start an evolution rather than an extinction. 1,000+ musicians "co-write" a silent album to protest UK's AI proposal. 92% of UK undergrads use AI for their studies. OpenAI expands Deep Research to more ChatGPT users. Sakana AI walks back claims about its AI CUDA Engineer. AI-generated "inspo" is warping customers' sense of reality. Alibaba to invest $53B on AI infrastructure. US schools adopt AI "wellbeing companion" for mental health support. Journalists take freelance jobs with AI training data companies. AI video of Trump and Elon blasted on government office TVs. Meta showcases its latest AR glasses, Aria Gen 2. 20,000+ once-public GitHub repos still accessible via Copilot. Microsoft cancels leases for US data centers. Gemini says SEC closed its investigation without action. Claude plays Pokémon on Twitch. Jürgen Schmidhuber discusses Saudi Arabia's AI ambitions. Meta in talks to build $200B AI data center. Perplexity teases Comet browser for "agentic search". ECMWF's new weather model outperforms others up to 15 days ahead. AI boosts productivity in SV startups, fueling "tiny team" success stories. Activision confirms using AI in Call of Duty: Black Ops 6. Chegg sues Google, saying AI Overviews hurt its traffic and revenue. Alibaba's open video model becomes AI porn machine. Slop publisher sells ripoff book on Amazon. OpenAI expects revenue to triple in 2025. Mira Murati's startup aims to raise $1B. Amazon values Anthropic stake at $14B. Blogging in the age of AI is still worth it.
Have you had a chance to play w/4.5 yet? I have to wait for next week.