🤝 Today’s edition is brought to you by AIOZ.

AIOZ AI’s first AI Challenge is complete, and the future of Web3 AI is wide open. Explore monthly challenges on the Web3 collaborative marketplace, powered by AIOZ Network DePIN.

📰 Welcome back!

For those of you old enough, hark your minds back to the March 2000, when there was a similar conversation with the growth of the internet.

“When is the ‘.com bubble’ going to burst?” they said. And just like that, share prices plunges while plenty of other collapses followed resulting in trillions of dollars evaporating into thin air.

Now, in 2025, the same question is being asked again, only this time, it’s about the AI boom. Is history going to repeat itself? Or is it different this time round? We’ll break that down and plenty more in this week’s edition of the Big Machines newsletter.

🚀 What we’re covering today…

• 🫧 Is the AI bubble going to burst?

• 📱 Google doubles down on AI phones with Pixel 10 series

• 💬 Thousands of Grok chats are leaked on Google

• 🧬 Protein folding secrets revealed using GPT

• 🤓 GPT5 is a fucking maths nerd

• 💲 DeepSeek V3.1 cost comparison is absolutely nuts

• 🤐 Did Musk court Zuck over an OpenAI takeover?

• 💧 How much water an AI Google prompt really uses

• 🚀 Microsoft says it’s ‘dangerous’ to study AI consciousness

🔴 Quick Note: We like to cover loads of AI news in our newsletter, so for a better reading experience, we suggest opening this in your browser for the full experience! 

Head to the ‘READ ONLINE’ tab at the top of this email.

👁️ 👁️ What you might have missed

• Is the AI bubble really about to burst? Well, according to some of those on Wall Street, they fear that same popping sound that haunted the internet in early 2000 could be right around the corner for those embedded in the world of artificial intelligence. On Tuesday, tech stocks suffered a huge shock sell-off after a report from researchers at Massachusetts Institute of Technology (MIT) warned that many AI investments were wielding a “zero return” for many businesses. Shares in $4trillion company Nvidia have dropped by 3.5pc while Palantir’s stocked dropped nearly three times as much. Since ChatGPT’s launch in 2022, Silicon Valley has been adamant that chatbots will lead the AI revolution and boost the economy for US companies. But that revolution has slowed down, according to MIT researchers and now the bubble is about to burst. Are investors getting overexcited by the AI boom in recent years? Probably, but can you blame them? Others believe that this setback is not a brick wall, more a speed bump that will slow things down, rather than bring it to a complete stop. Nvidia are set to report its results and could be a telling sign of what the actual state of AI is currently. But until then, keep an ear out for that dreaded popping sound.

• The launch of the Google Pixel 9 series last year saw the integration of AI and it became a prominent feature, with features like Gemini Live, image-generation tools and searchable screenshots. Now, with the launch of the Pixel 10, Google is rolling out a fuckload more AI features as they look to create distance between themselves and the likes of Apple. These include their Visual Overlays feature for the camera, Voice Translate for calls, Camera Coach, a proactive “Magic Cue” feature and loads more. For example, the Camera Coach feature, built into the camera app, will ask you to scan the scene you want to photograph. Then it’ll ask what you want to include in the photo while offering different example shots based on what you’ve asked it. So now you’re no longer an ‘Instagram photographer’, with the help of AI, you’re a real, professional snapper! So now if you’re filming an orgy and don’t want to include the really old guy in the shot, Camera Coach will be there to assist you. (Yuck.) Moving on to Magic Cue, which will help suggest replies in messaging apps using personal information from your Google apps to help respond in the most helpful and invasive way. There are loads more features, so watch the rundown below from 9to5Google:

• Grok’s fucked up, again. What is it this time, you ask? Well, hundreds of thousands of conversations users have had using Grok are now easily accessible on Google Search. ChatGPT users were met with the same fate only few weeks ago when thousands of chats were leaked online, sparking fury over confidentiality concerns. Now the same thing has happened with Grok with users hitting the “share” button on chats seeing their conversations with the chatbot created in its own unique URL. However, according to Forbes, these URLs were being indexed on search engines like Google and Bing, allowing anyone to view Grok users weird and fucked up conversations, including instructions on how to cook meth and make fentanyl, tips on how to make bombs, and even a detailed plan for how to assassinate Elon Musk. Now even Elon’s own chatbot wants him dead.

• A few Big Machines newsletters ago, we reported which jobs would be likely most affected by AI. While there were some surprising entries on the list, the gist was that AI would be coming for most jobs you could think of. And it appears over 70% of Americans fear that this will become a reality in a recent poll. However, quite surprisingly, the top concern wasn’t that AI would take American jobs - that was second. What was the most concerning thing AI could influence was “political chaos caused by US rivals”, according 77% of the 4,446 respondents. Funnily enough, it seems like Trump is doing that exact thing without using AI at all.

• Gym bros, get ready to get hype: Researchers at MIT have been using AI models to predict which proteins that might make good drug or vaccine targets. Now while this sin’t quite a full chicken at Nando’s protein gym bros can’t get enough of, it’s frankly more important. The models, which are based on LLMs, can make extremely accurate predictions of a protein’s suitability for a given application. MIT researchers are now using a novel technique to open up that “black box” and help them determine what features a protein language model takes into account when making predictions. Essentially, understanding what happens in that black box will help streamline the process of identifying new drugs or vaccine targets. Pretty cool.

• ChatGPT-5 may have had a slow start at getting off the ground, but some users have found that it’s wicked smart at mathematics. One user on X put forward the claim that GPT-5-Pro can prove new interesting mathematics. With proof, he then asked GPT-5-Pro to work on a clean open problem from a convex optimisation paper. GPT-5-Pro proved a better bound than what was in the paper, but noted that there was still some progress to be made. Loads of maths jargon I just simply do not have the brain capacity for, but if you do, check it out in the post below:

• DeepSeek-V3.1 was revealed to masses last week with its hybrid inference, faster thinking and stronger agent skills all promising enough updates for users to sink their teeth into. But we’re not here to talk about that. We’re here to discuss how madly cheap it is, compared to ChatGPT 4.5; and there’s a bar chart to visualise how insane it is. In the post below, the first bar chart shows how DeepSeek’s newest model performance is much better than that of GPT-4.5-preview, so it’s already outperforming OpenAI’s chatbot on that front. But it’s the price comparison in the second graph that is truly nuts. DeepSeek’s newest model costs just $0.99, a whopping $182 less than GPT-4.5-preview.

• Elon Musk has been courting Mark Zuckerberg in an attempt to acquire OpenAI, according to the OpenAI’s lawyers. They company are now asking Meta to produce evidence related to any coordinated plans with Musk and xAI to invest in the ChatGPT-maker. This has all come about due to Musk’s ongoing lawsuit against OpenAI. Their lawyers have revealed they subpoenaed Meta in June over Musk’s potential involvement in an unsolicited $97billion bid to takeover OpenAI earlier in February. And now there are claims from OpenAI’s legal team that Musk has been communicating with the Meta CEO about a bid to purchase the AI giant. Juicy. We first reported on this story in one of our very first newsletters and told you to watch this space with this story. It doesn’t seem like it’s anywhere close to being settled, either.

• How many drops of water is used to generate a prompt text from Google’s Gemini AI assistant? That’s been one of the key environmental questions that has been debated recently, with Google claiming it’s only five drops per prompt with the study revolving around the amount of water that data centers use in cooling systems to keep servers from overheating. However, according to experts, Google’s being a little bit misleading with those claims. The tech giants estimates that they’re only using up around 0.26 millilitres of water, or about as much electricity if you were watching TV for nine seconds, for every text prompt you send. Their estimates are down on previous research paper, which they owe to the fact their way more efficient. However, experts told the Verge that Google left out many key data points in their study and they’ve been fibbing out critical information, especially how Google have omitted indirect water use in its estimates. They also left out power consumption and pollution in their study which only skews the numbers even more.

• The topic of “AI welfare” has been a raging debate in Silicon Valley, with some believing AI models that might develop subjective experiences to living beings and “consciousness” should then be given some rights. It sounds a bit fucking mental and that’s a stance Microsoft’s CEO of AI, Mustafa Suleyman is taking, saying that the study of AI welfare is “both premature, and frankly dangerous.” While AI could develop consciousness in the future, the fact researchers are exacerbating human problems that are becoming prevalent in society, with humans having psychotic breaks and weird attachments to chatbots at this early stage of AI’s existence. While it’s a pretty level-headed take, that we’re building AI for people, not to be a person, there are many in the industry that are against this idea, with Anthropic and OpenAI both embracing the idea of studying AI welfare. It’s all a bit fucking weird, innit?

🧩 Other Bits

• It’s a new dawn, it’s a New Delhi for OpenAI, who have just announced plans to open its first office in India – just days after launching a ChatGPT plan tailored specifically for Indian users. The AI giant is looking to tap into the country’s rapidly growing AI market and is set to open up shop in the country’s capital in the coming months.

• China have just dropped their answer for Cursor: Qoder in its unveiling took aim at all those AI tools that ‘nail the demo’ but then completely shit the bed on the actual codebase. The next-gen Agentic Coding Platform promises to understand your who codebase, patters and even ships code that fits. Early results are looking promising.

• Is GPT-5 already old news? Sam Altman is already talking about GPT-6 (just days after the newest release) and people’s desire to need more memory. Altman also wants future versions to let users define tone and personality and says a newer update will arrive sooner than the gap between GPT-4 and GPT-5. Tell me GPT-5 isn’t landing without telling me it isn’t really landing.

• Things are about to go nuclear at Google. The tech giant are now one step closer to achieving its nuclear ambitions after inking a deal to work with public power utility Tennessee Vally Authority (TVA) to buy electricity from a next-gen reactor. The deal is the first of its kind and the plan is to start supplying juice to the local grid that serves the Google data centers in that area.

🤝 In Partnership with AIOZ

The inaugural AIOZ AI Challenge has wrapped, and judging is underway! 🎉 Congrats to all 650+ submissions for experiencing the collaborative power of the AIOZ AI marketplace.

Did you know? The AIOZ AI Challenge feature lets anyone host competitions directly on the platform. You can explore submitted models, track the leaderboard, or even launch your own challenge.

This is just the beginning. Join the movement to build, share, and monetize with AIOZ AI.

📈 Trending tools, models & apps this week

📋 LLM Leaderboard

📲 Trending tools & apps

🫵 Our Picks

• Mirage 2 uses AI to take any image, from a photo to a kid’s doodle, and instantly transform it into a live, explorable 3D world. You can then tweak the scene with text prompts, add surreal events, and share a link so anyone can jump in and experience it with you.

• Qoder is an AI-powered coding platform that reads your entire codebase, understands your patterns, and writes production-ready code that actually fits — even in messy legacy projects. This is Asia’s answer to Cursor!

• Google’s Veo 3 its latest high-end video generation model is rolling out to free Gemini users tomorrow, giving everyone a taste of pro-level AI filmmaking. If you haven’t used it yet, go try it out; it is a big zero-to-one moment.

🤓 Educational Picks

• Dex uses AI-powered image recognition to help kids (ages 3–8) learn new languages by snapping objects in the real world, with interactive stories and games that turn real-life moments into fun lessons.

• ChatGPT’s new flashcard quizzes let you turn any topic into interactive study cards on the spot, making learning sessions more engaging and memorable.

• Anthropic has released three free AI fluency courses, co-created with educators, to help teachers and students build practical and responsible AI skills.

🚀 Trending Apps & Models

• DINOv3 Video Tracking uses Meta’s latest self-supervised vision model to do semantic tracking directly in your browser, with no server needed, zero cost, and all powered by WebGPU and Transformers.js.

• MoDA Fast Talking Head blends a static image and your voice to create realistic talking avatars using a multi modal diffusion architecture, resulting in smoother lip sync, expressive head movements, and faster, more efficient rendering.

• AI Companionship Leaderboard scores how empathetic different AI models are based on emotional support, vulnerability, and relational traits, using the INTIMA benchmark to visualize which assistants are most or least human like in their responses.

💸 Financials

• Google have reportedly signed a $10billion six-year cloud deal with Meta, which will see Mark Zuckerberg’s minions use Google Cloud's servers, storage, networking and other services as part of the agreement. It’s the first major cloud partnership between the two companies as Zuck continues to scale his AI infrastructure as quickly as possible. On the other end of the deal, Google will distribute Meta’s Llama models.

• Another day passes and another day we stray further from God’s light. Lol only kidding. FieldAI have raised $405million to build universal robot brains. While they’re unlikely to be mushy, you can expect this California-based robotics startup to start pumping these bad boys out very soon, with the brains set to go into humanoids, quadrupeds and self-driving cars.

• Two Harvard dropouts are planning to really invade your privacy with their ‘always on’ AI smart glasses that will listen and record your every conversation. Taking a leaf out of former Harvard alumni Mark Zuckerberg, who is the king of taking your data, the two ex-students want to make you “super intelligent” and to give you “infinite memory” just by wearing them. It’ll be a hard pass from me, I’m afraid.

• SRE.ai has raised $7.2million for DevOps AI agents, with Salesforce Ventures and Crane Venture Partners leading the seed round. Raj Kadiyala and Edward Aryee, who both worked at Google Research and DeepMind, created SRE.ai to offer more modern tools to enterprises so they can avoid issues like metadata merge conflicts.

🕵️‍♂️ FREE ENTRY TO OUR INVITE-ONLY AI CHAT ON TELEGRAM…

If you share this newsletter with a friend and they actively sign up for the Big Machines newsletter, we will send you access to our invite-only Big Machines Telegram group, which is full of builders, investors, founders, and creators.

Access is now only granted to those who refer our newsletter to active subscribers, which means if you sign up on your work email, we will know you sneaky bastards.

This would kill our open rate, so please don't do that, we beg.

👋 Here’s some memes to get you through the week… see you next time

Here’s another reminder not to put your full trust in what these chatbots send back to you when you give them prompts.

On reddit, one user was asked by ChatGPT if they wanted a diagram of what’s going on inside their pregnant belly, with the user confirming it’s a boy.

The results are frankly incredible.

Finally, some more AGI from GPT5… we don’t care if it is fake…

That’s all for this week. In a while, paedophile crocodile!

Matt, Sam, Grant, Mike and The Big Machines team.

Follow us on Twitter.