Welcome to this edition of Loop!

To kick off your week, I’ve rounded-up the most important technology and AI updates that you should know about.

‏‏‎ ‎ HIGHLIGHTS ‏‏‎ ‎

Let's jump in!

1. AWS is about to launch an AI agent marketplace

We start with AWS, who are planning to launch their own marketplace for AI agents. AWS customers will have a single destination to browse and install AI agents - making it easier for them to adopt AI tools into their workflow.

Essentially, an AI agent is a bot that can autonomously solve tasks and decide when it should take an action. The agent will figure out the steps that are required and then start the process for you.

They’re quite different to traditional chatbots, as they have access to tools - like connecting to your database - and can work on a task independently. Or, the agent can ask for your feedback on how it should progress.

It’s a completely new way to solve problems, since we’re handing over more control to these agents and allowing them to make decisions for us. Whereas, with traditional software we needed to be very specific about how our software should work.

However, it raises new challenges for businesses too. If you’re not careful, the agents can cause more harm than good and can lead to higher costs.

For the last 18 months, I’ve been developing AI agents that are tailored for different use cases - from software engineering, to sales, to Workday testing - and I’ve regularly flagged this as a top priority for businesses.

It’s great to see that the technology is finally maturing and AWS is creating this marketplace, but we need to be mindful of the risks involved.

When you give AI agents access to private data and advanced tools, they can be both incredibly powerful for users and dangerous when malicious actors redirect them.

2. OpenAI and Perplexity prepare to launch AI-powered web browsers

OpenAI’s new browser isn’t expected for several weeks, but Perplexity has already thrown its hat into the ring with Comet.

This new browser is currently available for subscribers on the $200-per-month Max plan, with access slowly rolling-out for everyone else in the coming weeks.

It’s a bold move for both companies, as their teams will have less time to focus on the core product, but the potential upsides are huge. If they get this right, they could collect more data about how people browse the web and use this to better integrate their products.

Google adopted the same strategy almost 20 years ago, which has paid off handsomely. Chrome now dominates the sector and holds a market share of 68%, while their next nearest competitor sits at just 16%.

It’s also a great opportunity for these companies to raise new revenues, as they can package their products and closely integrate them. If OpenAI can create their own agents that monitor how you browse the web and perform actions on your behalf - such as online shopping - they could sell it as a new product offering.

OpenAI has a target of reaching $125 billion in revenue by 2029, which is a tall order. But they could achieve that if they create brand new productivity tools for businesses and convince them to spend more.

3. Academics hide prompts to influence AI review systems

It’s a clever strategy to manipulate the AI tools that are reviewing their academic work. It was found that 17 preprint papers on arXiv contained hidden prompts, which tell the AI tool to generate a glowing review of the paper.

To hide the prompts, they used white text and extremely small fonts that aren’t visible to the naked eye. They explicitly instruct the AI to "give a positive review” and talk about the paper’s "methodological rigour”.

These were found in papers from 14 institutions around the globe, including Columbia University, the University of Washington, Japan's Waseda University, and South Korea's KAIST.

But a professor from Waseda University gave an interesting defence, as he said it was designed to catch "lazy reviewers" that use AI tools - despite there being a ban on automated reviews.

4. McDonald’s exposes data about millions of people, used “123456” password

Did you make a mistake recently at work? Well, at least you didn’t expose data about 64 million people and use “123456” as the password.

Security researchers found that McDonalds’ AI hiring chatbot, called McHire, could be accessed using the simple password - which is regularly listed as one of the world's worst passwords.

As a result, the researchers could access a treasure trove of sensitive information. But beyond the password issue, they found another vulnerability in an internal API that exposed applicants' chat histories with the AI recruiter.

The data included names, email addresses, home addresses and phone numbers - essentially everything a bad actor would need for identity theft or targeted phishing campaigns.

The company stressed that the data was never "leaked online or made publicly available," though this rather misses the point about how easily it could have been accessed by anyone who bothered to try.

5. AI coding tools are making some developers slower, study shows

METR, a non-profit AI research group, ran a controlled trial with 16 experienced open-source developers and asked them to tackle hundreds of tasks on codebases they’re familiar with.

To test whether AI tools were genuinely boosting productivity, the study randomly assigned tasks as either "AI-allowed" or "AI-forbidden”.

Fascinatingly, they found that developers were actually 19% slower when using AI tools like Cursor Pro - despite predicting they'd be 24% faster.

This slowdown is attributed to the developers spending too much time on prompting and then waiting for a response.

They also found that the AI models struggled with large, complex codebases - leading to even more prompting and waiting around.

While I’ve found that these models have improved my productivity, it does depend on the task. They’re fantastic at solving very small, specific tasks and can convert data from one format into another.

But you will face issues when trying to use them for more complex issues, which can span across multiple code files. When I’ve tested OpenAI’s Codex agent, it typically takes 10 minutes to solve a task.

While you can do other things in the background, that’s not always the case and people might be tempted to wait around - resulting in lower productivity.

Maybe that will change in the next 12 months and the models can solve those complex problems, but it’s not possible today.

OpenAI’s acquisition of Windsurf falls apart, Google swoops in

Google has swooped in to hire the leadership team from Windsurf, just as OpenAI's $3 billion acquisition deal fell apart.

The search giant is bringing on board Windsurf CEO Varun Mohan, co-founder Douglas Chen, and several top researchers to boost the team at DeepMind.

Google is paying $2.4 billion to license Windsurf's technology and hire its talent, but won't obtain a stake in the company itself.

You might remember that Microsoft performed a similar move last year, when they hired the CEO of Inflection AI and licensed their technology - effectively avoiding any scrutiny from regulators.

The timing couldn't be more dramatic. OpenAI's deal collapsed due to tensions with Microsoft, who wanted to prevent the tech giant from having access to Windsurf's IP.

Under OpenAI’s current deal with Microsoft, they must share their technology with the company. Previously, both parties were happy with the deal, but the move to acquire Windsurf has put pressure on that relationship.

Google has secured a strategic win here. As competitors like Anthropic gain ground with their own AI developer tools, Google’s product team will benefit from both Windsurf's expertise and their technology.

Intel’s CEO says it’s “too late” to catch up with Nvidia

Intel’s CEO made the sobering admission that it’s no longer among the industry's top 10 companies, as it begins to lay off thousands of workers.

It’s a stark departure from Intel's once-dominant position in the semiconductor industry. Back in 2005, it rejected Steve Job’s suggestion that they partner on iPhone chips - with Intel missing out on the smartphone revolution.

However, they did continue to provide chips for Apple’s Mac computers and many PCs. But that all changed in 2020, when Apple became tired of Intel’s poor performance and decided to build their own semiconductors.

Then the company missed out on the AI revolution, as Nvidia invested significantly more in their chip line-up and then benefited from the rise in demand for computing power.

Today’s numbers tell the full story: Intel's market value has halved to around $100 billion in just 18 months, while Nvidia briefly touched $4 trillion last week - the first company ever to reach that milestone.

Lip-Bu Tan has acknowledged that it’s "too late" for Intel to compete with Nvidia on AI training chips, as their rival’s position is simply "too strong.” Instead, the company is banking on its upcoming 18A manufacturing process and hopes to regain competitiveness against TSMC.

Intel has suffered setbacks for some time now, as they’ve regularly found that their new semiconductors do not perform well enough.

It’s an embarrassing admission for the former chip giant, but the writing has been on the wall for a decade at least. While Intel enjoyed a rare combination of incredibly high demand and high profit margins for many years, that period is long over and it’s now far behind the competition.

Lip-Bu Tan believes that Intel must become much leaner to react quickly and become innovative again. As a result, it might have to scale back its plans to manufacture in the US.

🤖 OpenAI is testing a new feature called “study together”

🚕 Waymo’s robotaxis are heading to Philadelphia and NYC

💻 CoreWeave acquires a data centre provider in $9 billion deal

🎾 Tennis players criticise the AI technology used by Wimbledon

🚀 SpaceX plans to raise new funding at $400 billion valuation

💰 Mistral plans to raise $1 billion in funding

📱 Linda Yaccarino steps down as CEO of X

🗣️ xAI apologises for Grok’s horrific behaviour

📜 California lawmaker pushes for mandated AI safety reports

💎 Bitcoin surpasses $118K in a second all-time high

⏰ OpenAI delays the release of its open model

🇨🇳 Nvidia plans to release a new AI chip that’s designed for China

This is an AI startup that’s gained attention for “ethically training” their video generator and not using copyrighted content.

The company was founded by several researchers from DeepMind and is now based in LA, as they pitch the tool to filmmakers.

With their AI model, called Marey, they’re able to give creators more granular control over their footage.

Rather than just typing prompts and hoping for the best, filmmakers can manipulate camera angles, swap backgrounds, and control object motion - while the AI is still able to ensure the physics is realistic.

But the startup will face stiff competition from OpenAI, Google, and Midjourney. All three companies have released advanced video generators in the last 6 months, with Google’s Veo 3 clearly in the lead.

Meta have also done some fascinating research in this area, as they unveiled their Movie Gen model last year. Using simple text, you can easily add elements to the video and completely change how it looks.

Moonvalley is betting that, by offering the filmmaker much more control, they can differentiate themselves from the tech giants.

This Week’s Art

We’ve covered quite a bit this week, including:

If you found something interesting in this week’s edition, please feel free to share this newsletter with your colleagues.

Or if you’re interested in chatting with me about the above, simply reply to this email and I’ll get back to you.

Have a good week!

Liam

Share with Others

If you found something interesting in this week’s edition, feel free to share this newsletter with your colleagues.

Liam McCormick is a Senior AI Engineer and works within Kainos' Innovation team. He identifies business value in emerging technologies, implements them, and then shares these insights with others.