🖥️ Google and OpenAI are in a race to change the internet

Welcome to this edition of Loop!

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

‏‏‎ ‎ HIGHLIGHTS ‏‏‎ ‎

• Why the US Government could spend over $32 billion a year on AI

• How students can use Google’s newest AI tools for education

• OpenAI disbands their team to control super-powerful AI

• … and much more

Let's jump in!

1. US Government could spend over $32 billion a year on AI

It’s been a very busy week, but we start with a bipartisan working group in the US Senate and their recommendation that over $32 billion should be spent annually on AI projects.

The report argues that huge investment is needed to keep the US ahead of their international rivals.

They’ve outlined several ideas to solidify the US’ position:

• Cross-government AI R&D collaboration

• Fund AI hardware and software projects that are based in America

• Expand the National AI Research Resource program, which was recently started as a pilot scheme

• Launch "AI grand challenges" to drive innovation in difficult areas, like medicine

• Support AI readiness with the general public

• Ensure there’s better cybersecurity practices for elections

• Modernise the federal government's IT infrastructure

• Assess the biggest AI-enhanced threats and how to minimise them

It’s worth stressing that this report’s purpose was to broadly outline ideas on what the US Government should focus on. It’s unlikely that this will get the green light during an election year, especially with such a hefty price tag.

However, it hints at the scale of ambition from US lawmakers and the new programs that might be coming down the track.

2. OpenAI disbands its team to control super-intelligent AI, key researchers leave

Several researchers in OpenAI's Super-alignment team, including the co-lead Jan Leike, have resigned due to disagreements with the company's leadership.

The Super-alignment team was tasked with coming up with technical work, which would allow humans to control superintelligent AI. They aimed to complete the work within the next four years.

The team was promised 20% of OpenAI's compute resources, but were often denied a fraction of that amount - which greatly restricted the work they could do.

Leike, who was involved in the development of ChatGPT, GPT-4, and InstructGPT, wrote on X to outline why he was leaving the company and his frustrations with the lack of funding.

Ilya Sutskever, an OpenAI co-founder who was instrumental in (briefly) ousting Sam Altman, was also part of this team - which is likely to have contributed to the internal friction.

Sutskever has also left OpenAI, almost a decade after he co-founded it with Elon Musk and several other AI researchers.

3. Slack faces criticism from users over their AI training policy

Customers expressed alarm online that their conversation history could be used to train AI models.

It came to light after a post on X highlighted Slack’s privacy policy, which stated that customer data is used to train their “global models”. If you want to opt-out, users were told to write an email to Slack’s support staff.

The company has since clarified that they don’t train generative AI models on the data, but the damage has already been done.

This case really highlights the unease within the technology industry and how big companies are using data without explicit permission.

There’s real anger over how tech companies are behaving, especially when it comes to AI and how these models are being trained.

Slack has found that out the hard way, but others will need to be crystal-clear about how they collect data. Otherwise, they could face a similar backlash.

4. Google’s GenAI models can now analyse hours of video

There were a lot of great announcements by Google at their I/O conference, which I cover in the next section. However, this has one of the biggest impacts out of them all.

Gemini 1.5 Pro is now able to analyse up to 2 million tokens, which is double what they announced just a few months ago.

For context, Anthropic’s Claude can analyse 1 million tokens - while OpenAI’s models can only consume 128,000 tokens.

This is a big deal, as we can now ask these models to analyse huge documents - such as detailed reports, which contain hundreds of pages.

It can also answer questions about videos up to 2 hours long, which will be incredibly useful for academic lectures and YouTube tutorials.

While these models will still make mistakes and hallucinate, it’s hoped that this will become less of an issue as time goes on.

If you want to get access to the 2 million token version, you can join Google’s waitlist.

5. Google adds Gemini to its education software

When it comes to Generative AI, a lot of the conversation revolves around programming and how it can boost office tasks.

But there’s a real opportunity around how the technology is used for education and helping students prepare for exams. Google is well-placed to tackle this, given that many schools have been using their Classroom platform to assign homework and virtually deliver classes.

Gemini has a range of new features that are focused on education, such as step-by-step explanations for a given topic, being able to create practice materials for tests, and providing access to data from different sources - such as Rice University's textbooks.

While teachers can use Gemini to create lesson plan templates, summarise research, and more quickly provide feedback to students.

Google has also released LearnLM, which is a family of AI models that are specialised for education tasks. They’re already being used in some of Google’s products.

And if that’s not all, you can watch an academic video on YouTube and ask their AI to generate quiz questions about that topic.

Again, it’s important that students and teachers check the answers from the AI - since there’s the risk of hallucinating. Students should learn how to use these tools effectively, as it will give them an advantage over those who don’t.

Google reveals a flurry of new AI products

Google kicked off its annual developer conference with a long list of AI product announcements - ranging from email, to AI video generation, to identifying when a scammer is ringing you.

Most of us use Google products every day, whether it’s search, Gmail, YouTube, Android OS, or are viewing ads online.

The vast majority of people won’t notice if a feature uses AI or not, and for the most part they simply don’t care. But Google is in a strong position to further change how we interact online, simply because we use so many of their services on a daily basis.

One demo that has gone under the radar is Veo, their new video generator that looks very impressive. In many ways, it seems to match OpenAI’s Sora when it comes to realism.

For businesses, this opens up new doors for what we can make. I do wonder if it could be beneficial for self-driving cars, which haven’t improved much in the last decade.

We’ve seen how AI-generated videos can help to train robots, maybe the same is true for autonomous vehicles. Time will tell.

However, Google’s biggest announcement was Project Astra - which is a powerful new AI assistant that can see the world around you, listen to your voice commands, and respond quickly to questions.

It’s still being worked on, but Google’s team showed how you could interact with it using your phone. Your phone’s camera allows Astra to see what you’re looking at and can use that to answer questions.

Staff were moving the camera around the room, asking it what it saw and then questioned how those items could be used. If you’re at home, you could be looking at some food in a fridge and ask “What could I make for lunch?”.

Their demo was absolutely fascinating and it should have grabbed headlines all week. However, OpenAI unveiled GPT-4o as their own competitor - just one day before Google did.

Of course, this was done on purpose to steal some of Google’s thunder and stop their rival from getting all the press attention. There’s certainly no love lost between the two companies, as each tries to out manoeuvre the other in the AI race and compete for talent.

OpenAI reveals their breakthrough GPT-4o model

Setting aside that GPT-4o has a terrible name (surely the 4S would be a more amusing option?), it’s another breakthrough from OpenAI that follows on from Sora (generates videos) and GPT-4 (generates text).

GPT-4o stands for “omni” which means it can analyse text, speech, and video content. In many ways, it’s similar to Google’s Astra assistant, as it can also see the world around you and answer questions.

It’s really quick when answering a user’s questions, which means a lot more use cases are feasible - such as real-time translations between people. They’ve also showcased how the model could be used as a tutor for students, or help you prepare for a job interview.

Another benefit is that it’s half the price of GPT-4 Turbo and is much faster at responding. Again, this really helps businesses move from a simple POC and actually develop a product that’s feasible.

Although I do have some concerns. My biggest problem is how this technology is used and secured. OpenAI mentioned that you can give ChatGPT access to your computer and show it what’s on the screen.

But many businesses will be terrified by this. Given that staff often work with confidential data about their customers and business strategies, it’s a nightmare waiting to happen.

This is compounded by the EU’s GDPR law and the mega-fines that companies face if they misuse customer data.

Giving ChatGPT access to their screen could help automate tasks, but it’s not clear how companies can restrict it from seeing their private data. We will have to wait and see how this plays out.

As with every other Generative AI model, GPT-4o will make mistakes, but the upside is that we are getting closer to a “one-model-does-everything” future. Previously, developers had to come up with complicated systems to try and do that, which rarely worked well.

Thanks to GPT-4o and Astra, that doesn’t look very far away.

🇬🇧 UK releases its own platform to test AI models for safety

💰 EU warns it could fine Microsoft billions for not providing info about GenAI risks

🧑‍💻 OpenAI signs deal with Reddit to use their data

🎵 Sony warns over 700 AI companies to not use its copyrighted music

📢 Netflix is building its own ad server

☁️ Matt Garman takes over as CEO of AWS

📸 Anthropic hires Instagram’s co-founder as head of product

💾 Qualcomm is working on an Arm-based AI server

🤖 Cruise founder Kyle Vogt is back with a robot startup

📱 Apple are planning a “significantly thinner” iPhone for next year, in major redesign

🚗 Tesla Supercharger chief was fired for challenging Elon Musk

💬 Anthropic release an AI tool that generates prompts for you

LanceDB

LanceDB is a startup based in San Francisco that builds databases for multi-modal AI. They already work with the biggest names in the industry, such as Midjourney, Character.ai, and Airtable.

It was founded in 2022 by Chang She and Lei Xu, who realised that traditional data infrastructure was making it difficult for teams to bring their AI models into production.

Their product is a vector database, which is able to encode the meaning of unstructured data - such as images, text, and videos.

It uses a custom data structure, known as Lance Format, which is what sets them apart from their competitors and allows them to handle petabytes of data.

They’ve recently raised $8 million in funding, which brings their total up to $11 million so far. To broaden their list of customers, the team is focusing on their open-source project - with the aim of upselling and providing more controls to enterprise users.

LanceDB is already seeing 600,000 downloads a month, which is really strong given that multi-modal adoption from enterprise customers is still quite early. It’ll be interesting to see what new features they come up with in the next few months!

This Week’s Art

While the week was completely dominated by GPT-4o and Google’s Astra, we’ve covered quite a lot in this week’s analysis - including:

• Why the US Government could spend over $32 billion a year on AI

• OpenAI’s super-alignment team is disbanded

• Slack’s controversy over how they train AI systems on customer data

• Google’s GenAI models that can now analyse hours of video

• Google’s new education tools and how students can use them

• How Google is positioning itself against OpenAI and Microsoft

• OpenAI’s GPT-4o breakthrough

• And how LanceDB is building databases for multi-modal AI

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.

About the Author

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