
You’re receiving this email because you subscribed to the newsletter. If you’d like to unsubscribe, you can manage your preferences here with one click (or use the button at the bottom of the email).
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
How "workslop" is becoming a productivity killer in the workplace
OpenAI's launch of Sora 2 and its controversial social media platform
DeepSeek's new model that could slash API costs by 50%
… and much more
Let's jump in!


1. Anthropic takes a step forward with Claude Sonnet 4.5
We start this week with Anthropic, which has updated their Claude Sonnet model. Sonnet sits in the middle of their lineup, with Opus being the most powerful (and most expensive to use).
Sonnet 4.5 is a good improvement on version 4, which was already seen as the best AI model for coding and general tasks. With 4.5, Anthropic has made a step forward - but not a quantum leap.
I’ve used the model for the last week and noticed that it is better able to pick up on nuances, compared to previous models. It also seems to be slightly better at writing and summarising content, but I strongly recommend that you use Opus instead for those kinds of tasks - as it has the extra intelligence that’s needed.
Personally, I prefer Claude over ChatGPT as its responses sound more human and go into much greater detail. If you haven’t used it yourself yet, it’s definitely worth trying. OpenAI has some great models, but you might find that Claude is better for both your personal and professional work.
Anthropic have also made their Claude Agent SDK available for developers, which they use internally to power Claude Code - an incredibly powerful tool that allows software engineers to accelerate their work. Teams can now use the SDK to create their own advanced AI agents, similar to Claude Code, and allow those agents to perform complex actions.
For example, you could create an AI agent that takes in a user’s request, breaks it down into smaller tasks, and then uses your own software tools - such as Workday, Slack, or Notion - to call APIs and complete those actions for you.
I’ve been working on the technology for the last two years and AI agents are incredibly powerful, but only when they’re implemented correctly. If you have a question about AI agents and potential use cases for them, just reply to this email and I’ll get back to you.

2. Ex-OpenAI researchers raise $300 million to automate science
Periodic Labs has emerged with a massive $300 million seed round, which has been backed by Andreessen Horowitz, Nvidia, and Jeff Bezos.
The startup's core mission is to automate scientific discovery and create AI scientists that are able to work together.
It was founded by a strong team who've led top research projects at both DeepMind and OpenAI. Ekin Dogus Cubuk previously led DeepMind's project to discover new materials with AI, while Liam Fedus is best known for his work on creating ChatGPT and served as OpenAI's VP of Research.
But the team aren’t just focused on creating virtual simulations, they’re also building physical laboratories where robots are able to run experiments and collect real-world data.
If they get this right, it could lead to entirely new datasets and fuel future advances, similar to those we’ve seen with LLMs and the huge amounts of data they’ve used online.
With such an impressive team - and experience from OpenAI, DeepMind, and Microsoft - it’s easy to see why the company has raised so much money from top investors.

3. California Governor signs a landmark AI safety bill
The new legislation will force California’s top technology companies to become more transparent about AI safety and provide protections for whistleblowers - potentially setting a national standard that other states might follow.
Companies will need to report safety incidents to California's Office of Emergency Services, which can include AI-enabled cyberattacks, deceptive models, and critical failures that could harm ordinary people.
As you’d expect, the industry has been deeply divided on the legislation. Anthropic naturally leans more towards safety and has backed the measure, while OpenAI and Meta have actively lobbied against it.
Those companies fear that fragmented, state-by-state regulations will hamper their innovation work and introduce new barriers.
OpenAI tried to stop the bill at the last-minute, with a letter sent directly to Governor Gavin Newsom that proposed an entirely different framework - a move that irritated California’s legislators.
It’s worth noting that Meta has also become more active politically in recent weeks, with it creating a super-PAC to lobby against new AI regulations.

4. OpenAI allows you to shop with ChatGPT
Following recent changes to ChatGPT, American users can now purchase from Etsy sellers directly within the chat interface - and over a million Shopify merchants will be able to do the same shortly.
The new "Instant Checkout" feature lets users tap a "Buy" button, without having to leave the chat. Depending on how you’ve set it up, your payment will be handled through Apple Pay, Google Pay, Stripe, or a credit card.
If this starts to catch on and people do use ChatGPT to buy more items online, it could cause real harm to Google and Amazon. Both of those companies have strong platforms that allow us to discover new products and buy them in seconds.
But if OpenAI handles this directly, then Google and Amazon will no longer see those transactions and won’t be able to collect valuable data for advertisers - eventually eating into their revenues.
OpenAI has also partnered with Stripe to develop an open-source Agentic Commerce Protocol, setting up a direct challenge to Google's own Agent Payments Protocol.

5. DeepSeek releases a new model that cuts API costs in half
The Chinese AI company has unveiled V3.2-exp, an experimental model that could slash running costs by up to 50%. The researchers have introduced DeepSeek Sparse Attention, which is a clever way to tackle one of AI's most expensive problems.
Traditional transformers will process the entire context window, but their new system uses a "lightning indexer" to identify the important text. Specific tokens are then taken from those excerpts, which allows DeepSeek to use significantly less computing power.
The timing couldn't be better. Companies like Replit and Cursor are watching their inference bills climb as users hammer these models with longer, more frequent requests.
These aren't trivial expenses either, as they can spiral into hundreds of millions rather quickly - which has already forced some startups to sell-up. Windsurf’s founders have since moved to Google and their remaining IP was sold to Cognition.
If you want to dig into the technical details, you can access the model on Hugging Face and I’ve shared a link to their research paper below.

OpenAI launches Sora 2 and its own TikTok-style app

If you’ve recently seen videos of Michael Jackson doing standup, or Spongebob being pulled over by the cops, you’ve already seen the impressive results from OpenAI’s new video generator.
The company has just released Sora 2, which is a significant leap forward from the previous version. I’ve been really impressed with the new model, as it’s able to create clips that look much more realistic and consistent - even when the camera pans away for a few seconds and then pans back again.
Sora is particularly good at handling physics, too. If a basketball player misses a shot, the ball actually bounces off the backboard rather than magically teleporting to the hoop.
When you compare it with Google’s Veo 3, there’s not much to choose between them. Both models look incredible and can generate audio that matches the video clip.
Veo 3 will likely remain the top choice for film making and visualising shots, before actually making them in the real-world. Whereas Sora seems to be better suited for short, viral clips.
Controversially, OpenAI has launched a social media app that looks suspiciously like TikTok. With this new app, you can easily view clips that were generated by AI and share them with others.
It also has a feature called "Cameos," which allows people to upload images of themselves and then star in their own for AI-generated videos. Essentially, it allows people to create deepfake videos of themselves and make them go viral online.
I have no doubt that young teenagers will love this and make it into a trend, similar to how they gravitated towards TikTok and made it a driving force in our culture.
But is that something we really want? After all the harm that social media has caused to young people, and wider society, this seems like a terrible vision of the future.
However, most people online have missed that this is a huge data opportunity for OpenAI. With this new social media platform, they will now be able to collect incredibly detailed data about who uses the platform and what their interests are.
For TikTok, it has been an absolute goldmine and can tell them about your age, gender, political views, interests, and buying habits.
Clearly, OpenAI wants to replicate this and use it to create detailed profiles for advertisers - alongside building their own ecosystem and keeping you engaged with their growing list of products.
With OpenAI planning to reach revenues of $100 billion by 2028, we’re starting to see their strategy for making that happen.

Are your colleagues creating “workslop” with AI?

The workplace has a new enemy, and researchers have given it a name: "workslop".
This new term, courtesy of BetterUp Labs and Stanford Social Media Lab, describes the AI-generated content that's flooding offices everywhere - something that looks professional on the surface but lacks any real substance.
Both organisations have surveyed 1,150 employees in the US, which has revealed the scale of the issue - with 40% of people reporting that they've received workslop from colleagues in the past month.
Researchers define workslop as AI-produced work that "masquerades as good work", but fails to actually move projects forward. It seems to be a growing phenomenon and might explain why 95% of organisations see no ROI with AI projects.
But the problem isn't just that workslop exists, it's that it can create a domino effect of inefficiency throughout your business.
When someone submits AI-generated work that's incomplete or missing important info, they're essentially passing the buck to their colleagues - who then have to waste time interpreting, correcting, or completely redoing the task.
While the first person was able to generate that work in just a few seconds, thanks to tools like ChatGPT, it shifts the burden of work downstream and turns a productivity gain into a net negative.
I think we have all seen this at some point, especially in the last year as AI tools become more advanced and therefore open up new use cases. I regularly flag the valuable ways it can be used, but like any tool it can be abused.
In my view, the best way to stop this is for company leadership to remind people about what’s expected and what’s not. It’s important that we use these tools to accelerate our work and improve the quality, but that’s not possible if people blindly accept the AI tool’s response - without actually checking it and making their own changes.
Otherwise, we’ll be spending more time correcting work than actually moving forward. These AI tools are fascinating and can act as an expert over-our-shoulder, but they can also tank productivity when people rely on them too much.

🎵 Universal and Warner Music close in on AI licensing deals
🖥️ Intel is in talks to handle some of AMD's chipmaking
💰 OpenAI is now the world's most valuable private company
🏥 Google releases research about their health agent
👓 Apple shelves Vision Pro overhaul, will focus on AI glasses instead
🎬 Hollywood celebs outraged by the AI-generated actress Tilly Norwood
🛒 Amazon is launching a grocery brand, most products will be under $5
💬 Meta plans to use your AI chat data and sell targeted ads
📚 Wikipedia makes it easier for AI agents to access its data
🤖 Samsung partners with OpenAI to deliver chips for Stargate
🐕 Ring cameras can now find your lost pets
🚗 Zoox expands into Washington DC with its autonomous vehicles
🩸 Whoop expands into blood testing, now available to 350,000 on waitlist
👔 AI hiring startup Alex secures $17 million to automate job interviews
🔒 OpenAI rolls out parental controls on ChatGPT



Einride
This Swedish company is working to create self-driving trucks and has just raised $100 million from investors.
The startup is based in Stockholm and their fleet includes electric trucks, autonomous pods that look like something from the future, and they also develop software for logistics companies.
Einride's electric lorries are now used by PepsiCo and Carlsberg to move their products across Europe, North America, and the UAE. But their autonomous pods often attract attention for their distinctive looks.
They were designed to work without any steering wheels or pedals, with Apotea and GE Appliances already using them to move goods in America. They’re significantly smaller than a normal European lorry, but they could be useful for moving smaller goods - like parcels.
It’s worth noting that Einride isn't the only player here. Aurora is a US company that’s also developing autonomous trucks, which are able to drive at night and have already logged 20,000 driverless miles.
It’s an incredibly expensive business that’s usually reserved for big players like Google (who own Waymo) or Amazon (who own Zoox). Even companies like GM have struggled with self-driving vehicles, as they were losing billions with their Cruise brand.
The technology has quite a bit further to go, as it has more scenarios to prepare for when compared to a robotaxi. With the huge costs involved, and the added safety risks when carrying heavy goods, this is likely why the company has opted for the smaller pod design.
This Week’s Art

Loop via OpenAI’s image generator

We’ve covered quite a bit this week, including:
Why Anthropic's Claude Sonnet 4.5 is a step forward for everyone (especially developers)
The ex-OpenAI researchers that have secured $300 million to automate scientific discovery
California's law that forces tech giants to be more transparent about AI safety
OpenAI allows you to shop with ChatGPT
Why DeepSeek's new model could slash API costs by 50%
OpenAI's launch of Sora 2 and its controversial TikTok-style social media platform
How "workslop" is becoming a productivity killer in the workplace
And the Swedish startup Einride that’s raised $100 million for self-driving trucks
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
Feedback
How did we do this week?

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