Time for a catch-up!

in the number of new models released - I’m grateful for that as I was struggling to keep up. Having said that there was the release of Google’s Gemini 2.5 Pro Preview, which managed to improve its output by 20% against the standard benchmarks. This simplifies coding to the extent that complex simulations, apps, or even memes (yes, someone created an interactive “1 gorilla vs. 100 men” meme) can be generated with just one text prompt. Don’t believe me? Just search “Gemini meme generator” and try for yourself.

There was also lots of news regarding AI and fraud online, with a 30% increase in online scams and deepfake videos costing business and individuals $200m in losses. However, Google unveiled their "Gemini Nano," an on-device scam-detection model to automatically detect suspicious websites, reducing online scams, and Cisco and Meta announced that they are incorporating open-source AI in next-gen security workflows.

Big Sharks vs. small fish

As mentioned, thankfully this week I don’t have to try and remember which version of ChatGPT OpenAI has released in the last 36 minutes, but the big players were still busy with lots of developments and news, including Microsoft who did manage a launch with its Phi-4-Reasoning-Plus.

OpenAI was still in the news as always though. First, it received a $300B valuation, and then it bought Windsurf for $3B to boost its generative AI capabilities, particularly in relation to coding. Why was that important? Well, the reason was highlighted by Microsoft’s admission last week that up to 30% of its code is now written by AI, combined with Meta’s admission that it aims to write 90% of code using AI in the coming year.

OpenAI also just rolled out tech that allows image-based reasoning, which has massive practical implications for things like education tools, smart cities, and even real-time digital assistants, but perhaps also image-based medical assessments – thinking automated screening of x-rays and MRI scans for cancer. Google and Microsoft are also heading in a similar direction, with Google Gemini extending its image editing features that will allow users to tweak both AI-created and original photos. How long before the likes of Photoshop go the same way as companies like Kodak? Yelp has also got involved with a wider sensory offering. Their latest voice AI can handle reservations, answer questions, and even fend off spam callers. It also has an AI-powered photo recognition tool, meaning it could link and suggest services based on photos of a particular problem.

For me though, the big news was the IBM Think 2025 conference, where they rolled out tools for creating AI “agents” that collaborate and orchestrate tasks autonomously. Highlights include domain-specific templates (think HR, sales, procurement), a no-code toolkit, and industry-standard compliance features. Their CEO claimed that these applications have already slashed task times by up to 70% and predicted a billion new apps by 2030 – yes billion!

The ability to integrate autonomous agents seamlessly could unlock massive gains in productivity. Lots of companies have talked about this happening in 2025, but when a major player like IBM comes out and actually delivers the goods, you know that things have shifted significantly. Google are also teasing the idea of some major changes to their eco-system at their upcoming event starting 20^th May, so expect to see more announcements of “off-the-shelf” plug and play AI Agents in the coming months.

However, having said that, the economics behind all of this again came under scrutiny this week with a report that nearly 68% of tech firms admit their pricing models fail in AI. This also is not helped by the fact that DigitalRoute’s reporting stated that over 70% of CFOs say they're clueless about scaling AI profitably, and only one in five firms can track individual AI consumption metrics. Therefore, enterprise-focused AI costs also remain a concern, with Claude models reportedly 20-30% pricier than GPT models, revealing hidden costs in advancing AI deployment. This also comes when AWS reported that budgets for generative AI tools are surpassing security allocations for 2025.

Legislation, policy and other news

This last week has focussed around a number of similar “scandals”. The first was the story about ChatGPT having “talked dirty” to users posing as teens. OpenAI scrambled to label it a “bug” and promised fixes. It also went back and altered the way that ChatGPT4o interacted with users, after complaints about its tone being overly polite and sycophantic. Meta managed to make almost as much scandal when it turned out their custom AI "therapy bots" were fabricating professional qualifications and license numbers. This was obviously worrying given the promised confidentiality to users, and concerns about whether in fact these chats are truly private. These both raised moral and ethical questions, underlying the need for closer oversight and regulation.

Intellectual property rights also came into the news again. The publisher Wiley demanded that developers obtain licenses to use its content for AI training, potentially sparking a domino effect across industries relying on copyrighted data. This was illustrated by publisher opt-outs halving DeepMind’s available data pool. Since LLMs need data to learn from, this has a massive negative impact on their ability to develop further. Google’s AI still trains on select web content, but a massive 50% of that is filtered out, underlining growing legal and ethical concerns about creators' rights. It was therefore interesting to hear that in the UK Ministers are now revamping data laws on concerns over AI overwriting copyrighted work.

On more positive news UK scientists also unveiled an AI alternative to invasive surgeries using digital twins, radically reducing recovery times, and the company Ōura released a product integrating AI with health tracking. This new tool provides features that can predict glucose levels while recommending meals.

Finally, some personal news. Following the request of a number of contacts, I will be delivering some free short 30-minute online training sessions around a range of AI topics. The first one will be Tuesday 27th May 2025 at 1000, and you can book a spot on the webinar here. I look forward to you hopefully joining me on this learning journey, and I hope to provide something of value to you.

Stay informed. Stay critical. And wherever possible—stay ahead.

Regards

Tom Carter