Relevant AI News - February 2025

Model Madness

February was model madness month. There seemed to be a new frontier model every few days!

First of all, there are still startups raising at very high valuations. Ex-OpenAI CTO Mira Murati is rumored to be raising $1B at a $9B valuation for Thinking Machines Labs, while Ilya Sutskever is also looking to raise at a $30B valuation for his company SuperIntelligence. This is a pre-product raise! But obviously very well known founders.

Amazon made some major moves, they unveiled a new Alexa+ which was LLM based and much more chatty. There has also been news buzzing about their new Trainium2 chips, which will be a recurring theme in this month’s update.

Anthropic unveiled Claude 3.7, its newest model with “thinking” capabilities. You can basically set a timer to increase inference time compute, which improves the results the longer you set it. This is in line with previous reasoning models like o1 and DeepSeek R1, which allows the model to map out what the possibilities are and give an answer that is more comprehensive and detailed. 3.7 was a real hit on Coding agents like Windsurf as well.

Google released Gemini 2.0 pro and flash, which beats out its previous models by a lot while also being very inexpensive to run. Google is still finding its niche but they were in fact the first to come out with Deep Research.

Nvidia unveiled earnings with revenue of $39.3B, up 12% QoQ but 78% YoY. Every time there is an Nvidia earnings event, the entire AI industry holds its breath. However, with more and more AI chips coming out, it may not be a full signal of the health of the industry. Surprisingly, the stock still dropped 8.5% the next day, but this could be more related to macro related events which have brought down the equity markets as a whole.

Microsoft released details of its quantum chip that claimed to have a release date in years, not decades. There seems to be cracks in the relationship between OpenAI, but I’ll let that unravel before making any assumptions.

Meta was strangely silent in February, but it did get into talks to acquire FuriosaAI. Is everyone trying to build their own chips now? They were also rumored to be building a massive $200B data center. Apple also announced a data center to power Apple Intelligence and 20k research jobs, it planned on spending $500B over 4 years. This is some serious capex.

OpenAI had a blowout month as well. They came out with Deep Research, copying Google’s name, and then later copied by Perplexity, which scans the internet for information before writing a prompt for the user. It is much more intuitive and user friendly than the other versions, but comes at a cost. It is $200 a month to get 120 queries and the plus tier gets 10 queries. There is a trend here that OpenAI is raising prices, either to justify its massive valuation with proper revenue growth, or to match very compute intensive tasks. As usual, an open source version came out only a few days later called Open Deep Research. Rumors are also that OpenAI is developing its own chips with Broadcom. 2026 appears to be the year that Nvidia could face some competition on inference market share with Amazon and OpenAI entering the market (along with smaller players like Groq and Cerebras). OpenAI also unveiled GPT-4.5, a non-reasoning model with the power of reasoning capabilities. With an extreme cost to run, we’re waiting in anticipation for GPT-4.5 mini or turbo. Sam Altman would also hint that GPT-5 would be a single interface where the user would not have to choose between modalities.

Speaking of Perplexity, they are shipping nonstop. They released Sonar, which is a search model that is hosted on Cerebras running Llama 3.3. Rather than comparing it to benchmarks, they said it beat other competitors by “user satisfaction”. They also announced Deep Research which I just mentioned above, and teased a new web browser called Comet. While Perplexity is shipping fast, it is almost to a point where I am not sure which product is successful and which is not. However, you could argue OpenAI is doing a similar strategy on trying to make a land grab on consumer surfaces.

xAI released Grok 3, which was the counter example to DeepSeek. With DeepSeek using cheap chips to train good models, xAI proved that spending a ton of money on GPUs works too. Elon also trolled OpenAI by offering $97B to buy the company to attempt to derail the conversion from non-profit to profit. Have to end with the response here which was hilarious: