On Monday, the full impact of the DeepSeek R1 model was felt as Nvidia stock dropped more than 15% in one day and the whole, “we just need more compute” paradigm was thrown into question. Deepseek sent the whole of the US tech sector into panic mode and everyone was waiting to see how they would respond.
[00:22.0] But before they could, Perplexity and many others raced to incorporate the DeepSeek R1 model into into their tools in record time as demand from users drove some impressively fast integrations. Then DeepSeek casually dropped another open source model, this time a multimodal one.
[00:37.7] The Janus Pro 7B was seen to be threatening to shake up the text-to-image market in the same way it had done to the text models. So far, though we haven’t heard much more about it. On Wednesday, the Financial Times reported that OpenAI had found evidence that DeepSeek used the US company’s proprietary models to train its own open source competitor.
[00:57.0] Many people have pointed out the irony of OpenAI complaining that someone has stolen their training data. Few people had much sympathy. By the next day, Microsoft announced that DeepSeek R1 was available on Azure AI. What a difference a day makes. ‘Thinking’ became the new big deal in AI.
[01:13.6] By the middle of the week, a ‘Thinking time’ clock appeared in Grok and hinted at what might be to come from Grok3. Grok3 does seem to be imminent. There were reports of it being seen ‘in the wild’. Although it’s not yet available to the public, it seems some users were able to get access to the model where the responses were shown as ‘sender: grok3’.
[01:32.1] A day later a new ‘Think’ button showed up in the ChatGPT interface, leading to a lot of speculation about the imminent release of o3-mini from OpenAI. Then on Friday, the fight back began when OpenAI dropped the o3-mini family of ChatGPT models. They introduced a standard o3-mini model and an unexpected o3-mini-high model.
[01:51.6] The early benchmark testing is suggesting that it’s comparable or better than the o1 models across maths, science and coding. This would seem to mark the start of a high intensity battle of models between all the top players, including the new kid on the block, Deepseek. It would seem that it is no longer acceptable to wait months to respond to a new leading model.
[02:11.5] AI companies will be expecting to one up each other in weeks. This is likely to lead to a massive acceleration in capability with everyone trying to hold onto the top spot, throw the half a trillion dollar Stargate project into the mix and we have the recipe for an incredibly explosive year outside of the Model Wars.
[02:27.8] There was a big ruling from the U.S. copyright Office when they stated that the use of artificial intelligence tools to assist in the creative process does not undermine the copyright of a work. This removes some of the ambiguity around these kinds of claims and clears a path for greater use of AI content in creative fields.
[02:44.8] Tesla announced that their Robotaxi service will be going live this June in Austin, Texas. The full self driving cars will have no steering wheel, no pedals and just two seats. Many people will be apprehensive about stepping into a vehicle without any ability to control it directly, but it seems like it won’t be long until this is a completely normal phenomenon.
[03:04.3] Elon Musk said that 2025 is “maybe the most important year in Tesla’s history”. Talking about the robo taxi service specifically, Elon said that “Tesla is putting our toe in the water gently at first, just to make sure everything’s cool”. To prove that Teslas can have a safety record “significantly above the average human driver” with Robotaxis rolling out and Optimus ramping up production, this really could be the biggest year ever for Tesla.
[03:29.8] As usual, there was a lot going on in the world of generative AI. Without any fanfare, Grok released a new Image Edit feature. Now, when you generate an image in Grok, you can use the ‘Edit with Grok’ button to request some edits and refinements to the image.
[03:47.4] Freepik has added the Pixverse 3.5 model to its suite. It specialises in creating videos in anime, cyberpunk, 3D animation, clay or comic styles.
[04:04.1] Krea announced that it’s added character consistency via the Hailuo model to its tools. Character consistency is such a roadblock for anyone creating multiple scenes, so these tools are crucial to success. LTX Studio added a new feature called ‘Composition’ to its AI tool.
[04:22.0] This feature allows users to use any image as the base for your final output. This allows far greater control when it comes to composing a shot and feels far more creative than simply relying on a prompt. Finally, a new music generation tool called ‘Riffusion’ has been released.