🖕 Why make one enemy when you can make two? That appears to be Sam Altman’s approach to business at the moment, as relations between OpenAI and Microsoft EVERYONE are going sour.

Tensions between the two companies are reaching boiling point with OpenAI executives pointing the finger at their major backers Microsoft over anticompetitive behaviour in the partnership. Oh, and Meta are catching strays from OpenAI, too.

We will also be covering the publishing of the OpenAI Files. It’s a juicy document that’s basically had all of the world’s biggest AI company’s emails read out loud and it doesn’t look particularly good for Altman.

But in better news, (our) Sam is BACK this week after the lesser-known McEvoy brother filled in for the last edition. I know, I know, I’ll be sticking around for the foreseeable so there’ll be no separation anxiety for a long while (until I’m away again next week…). Anyway, there’s plenty of news to get through, so let’s waste no time, shall we?

🗞️ What we are covering today…

• OpenAI starts fights with everyone

• The OpenAI Files – what’s just dropped and what it means

• The Trump administration makes another AI boo-boo

• World models are the new goal and the new holy grail

• Chinese livestream features TWO digital avatars

• A circus bear performs three consecutive backflips

• xAI are all gas and no permits

• And image generation has come a long way in three years (sort of)

🔴 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

• Earlier this week, Donald Trump’s plans about integrating AI into his government were leaked on GitHub, with the plans showing how Trump was preparing for the widespread use of AI in all government agencies. And it wasn’t just snippets of the document – it was the WHOLE document. The whole thing is too much to cover in a chunky paragraph on our newsletter, but the main points are that Trump wants an AI chatbot to take over administrative tasks, an AI analytics dashboard, which officials can use in real time, and a collaborations with Technology Transformation Services (TTS) under the leadership of Thomas Shedd. Not the sexiest of leaks but if the person who got them could do the Epstein Files and some Area 51 shit next, that would be great.
  

• World models are the new goal and the new holy grail for AI. That’s according to Stanford’s Fei-Fei Li and Meta’s Yann LeCun, who argue that language alone is not enough for AI to replicate human intelligence and understand our physical realm. They say there’s no spatial reasoning, memory or planning from conventional Language Learning Models (LLMs), so what do they suggest? Building ‘world models’ that map out 3D environments maps mentally for AI. The theory behind it would see AI agents able to perform complex tasks with a human-like approach and understanding, while boosting productivity across all industries with safer robots. There’s data challenges to pull something this monumental off but we can dream. Also, I hope the 3D model doesn’t map out my secret porn stash under my bed.

• China’s livestream boom just hit a major milestone this week with something special. Luo Yonghao, one of the countries top livestreamers, created not one, but TWO digital avatars of himself to host the stream, a new industry benchmark. Better yet, they were engaging with each other and the audience as a set of human co-hosts would do, reacting and talking while racking up 13 million views in the process. I imagine if I got two of myself to write this newsletter, they’d argue over mundane shit and it would never get published.

• While I’m no an advocate of the circus (Dumbo fucked me up as a kid), I am an advocate for what MiniMax unveiled this week. Posting on X, the company introduced Hailou 02, along with a pretty impressive video a number of circus acts, including a bear doing three consecutive backflips, a lion running through rings of fire and a clown horrifyingly turning to dust, all with the simple prompts that showed exactly what was being done in the clip. Encore!

• Talking of impressive AI-generated videos, the guys at Midjourney have outdone themselves again. They introduced their V1 Video Model this week, accompanied with a video showing off the model’s capabilities and it blew us away. Using the model, users can put still images in like you normally would and then press animate to make the magic happen, whether that’s with prompts or leaving it up to the AI to figure it out. It’s available to use now for $10 a month, and it’s being advertised as a tool that everyone could use.

• xAI are all gas and no permits, and are now facing a lawsuit as a result. Earlier this week, the AI giant received word that they will be sued after operating a fleet of natural gas turbines without permits just outside of Memphis. The gas turbines used have the potential to emit more than 2,000 tons of NOx per year, which is contributing to smog. The lawsuit alleges that xAI failed to obtain the permits that are required by both federal and local regulators. This all probably explains why Grok has been acting clunky in the last week.
  

• Mozilla has quietly started testing Perplexity AI as a built-in search engine, in its first move toward AI-powered search in the browser. While it’s not replacing the bog standard search engine, it is being used as an optional alternative and is still in the early stages. And when we say ‘quietly’, we mean they announced it on its Connect community forum, rather than a formal announcement on its blog or some sort of PR release. But, it might have to wait…
  

• Apple execs are rumoured to be plucking up the courage to put a bid in for the one and only Perplexity! Apple CEO Ellen DeGeneres Tim Cook may still have a fighting chance at keeping his position if they somehow manage to pull this one off… I would be very very very surprised if this amounts to anything, personally.

• Remember that dream catcher you bought while travelling round Thailand on your gap year, which now hangs in your bedroom near your Indian tapestry blanket? Well fucking bin that, there’s a device that now catches your dreams and then plays them back to you! Modem unveiled their ‘Dream Recorder’ device this week, an all-white, open-source alarm-clock looking bit of plastic that plays your dreams back to you as cinematic reels. It supposedly captures your subconscious in ultra-low definition, and then you just wake up and tell it what your dream was about (if you haven’t already forgotten). I hope it doesn’t record any of my sleep paralysis episodes, though.

• Had a go on the Ray-Ban Meta AI smart glasses yet? No, me neither. While they appear gimmicky and could become a staple fashion item in years to come, Meta are broadening their horizons still. That’s because Meta are now reportedly building AI smart glasses with Prada. It’s Mark Zuckerberg’s way into Italian high fashion, even though Meta has already collaborated with eyewear giant EssilorLuxottica. Meta even teased a collaboration with EssilorLuxottica brand Oakley this week, and those glasses could cost around $360. So with that, I dread to think how much the Prada smart glasses would cost.

🗣️ Other Titty Bits 

• Cluely just don’t miss. They've just dropped a marketing video about their team and while it doesn’t necessarily tell you exactly what they do, the attention-grabbing, super impressive clip speaks for itself.

• Fair play to Supabase. They’ve revealed that they’ve had more sign-ups last quarter than over the last four years. We’re sure a lot of that has to do with Lovable, which isn’t shocking to say the least.

• VEO 3 is now being rolled out. Google made the announcement this week and is available in 70 plus markets for AI Pro and Ultra subscribers.

• Want to make your mates look funny using image generation but can’t be bothered to switch from WhatsApp to ChatGPT’s native app? Well ChatGPT’s image generation is now available in the messaging app – and it’s available for everyone.

• Around 41% of Y Combinator startups are solving tasks that workers don’t actually need automated. Startups are chasing full automation where partnerships would work better.

• Taiwan has imposed technology export controls on Huawei and SMIC chips, dealing another major blow to the two companies trying to spearhead China’s attempts to develop cutting-edge advancements in AI chips.

• A new feature has been released for Grok called Tasks. You can now schedule tasks and Grok being the nice bot that it is will do it for you automatically and will send you a notification when its done. Good bot.

• Google have given us a glimpse into what the future looks like for Gemini and where they think AI is heading. With a bullet-pointed list, Google want to provide better reasoning, general agents and long context windows for their models, in what is a welcomed bit of news.

• Sending a voice note to your mate about how they’re making a giant lasagne under the pitch at Wembley Stadium, only to be interrupted by an advert? That is becoming a reality. Meta are now putting ads on WhatsApp in what could be an infuriating move for its 1.5 billion active users.

• Gemini 2.5 Pro and 2.5 Flash are now stable and generally available, which includes a preview of Gemini 2.5 Flash-lite, offering you a new cost-efficient model in the 2.5 series. Stable and generally available also sounds like my mental health status.

• Got a spare $200 lying around? Anysphere have just launched a monthly subscription plan, named Ultra, for their Cursor AI coding tool, which offers users 20x more usage on AI models from OpenAI, Anthropic, Google DeepMind and xAI compared to their $20-a-month Pro plan. If you do the maths, that’s like, I dunno, bang for your buck?

• Are we all set to become Barbie girls in a Barbie world? Well, as terrifying it sounds, we might just be. Mattel, the toy company behind Barbie, have teamed up with OpenAI to integrate AI into the dolls. This is exactly how we got M3GAN, guys.

• Science-backed culinary is coming soon. Kaikaku revealed their research preview of their 300-dimensional flavour similarity model created using ML. Graphs have never looked so tasty.

• Everyone needs a Berk Jernström in their life. The guy behind Tictail, which was bought out by Spotify, is now helping developers build one-person unicorns so they can achieve greatness too with Polar.

• It turns out AI and humans are not so different after all. That’s because Google DeepMind has discovered that its Gemini 2.5 Pro model has a full-blown panic attack when playing Pokemon, specifically when its Pokemon is close to death in battle. Same, hun.

• California is trying to regulate its AI giants… again. After failing to pass a bill that would’ve enforced companies to test their AI models for specific dangers last year, a new report has been published. It suggests a new framework is need that has more transparency and independent scrutiny of AI models. But could it get vetoed again and we just repeat this cycle? Maybe.

🖕OpenAI ditches ALL of its friends: Tensions reach boiling point with Microsoft as their deal turns sour... and Meta can get fucked too, apparently

OpenAI’s house of cards is falling down – and they’re actively burning bridges with it.

Earlier this week, news broke that OpenAI executives were considering accusing Microsoft, their major backer, of anticompetitive behaviour in the partnership.

To be fair, with Microsoft’s track record, it wouldn’t be a total surprise if it was true – just look at how Bill Gates treated Homer Simpson.

Tensions between the two companies is now at boiling point and their famed AI partnership could be coming to an end.

The Wall Street Journal reported that OpenAI wants to loosen Microsoft’s grip on its AI products and computing resources. Then, less than 24 hours later, it was revealed that OpenAI wanted Microsoft to own a 33% stake in the company, with a “reshaped unit” in exchange for foregoing its right to future profits. Lol.

Basically, Sam Altman told Microsoft to go fuck themselves.

Microsoft have been one of OpenAI’s biggest backers over the past three years, helping with the rise of ChatGPT, which has slowly embedded itself into our daily lives. Microsoft invested $1billion into OpenAI back in 2019 – and has since grown to include $10bn in total funding.

As a part of this historic deal, Microsoft are entitled to 49% of OpenAI Global LLC’s profits – and now Altman wants that number dialled the fuck down.

But who is to blame? Has Microsoft been anticompetitive with OpenAI or does Altman want to wriggle out of a deal that essentially help grow his company to lofty new heights?

OpenAI’s pending acquisition of Windsurf seems to have driven a wedge between them and Microsoft. Reports suggest that the anticompetitive behaviour accusations surround Microsoft insisting on full access to Windsurf’s intellectual property. However, on the other side of the coin, Microsoft reportedly feel a tad uneasy about OpenAI developing a competing Copilot product.

Days later after this all unfold, reports then suggested that Microsoft were ready to abandon high-stakes talks with OpenAI over the future of its alliance.

That’s because the tech giant didn’t want to take talks any further with OpenAI if they were unable to agree of the portion of the pie they’d be getting from their 11-year deal, which expires in 2030.

It’s far from over between the two companies, but by the looks of it, it could be a messy finish.

And Microsoft weren’t the only company Altman was taking jabs at this week.

OpenAI are now phasing out its work with Scale AI and unfriending them on Facebook too following the data provider’s deal with Meta.

It sounds like Altman doesn’t want to be seen dead with these companies and wants to go in alone, despite Sarah Friar, the chief financial officer of OpenAI, previously stating they wanted to continue working with Scale AI.

That tone has now changed and Altman is creating this ruthless cut-throat businessman image that I guess comes with the territory.

🏝️ The Epstein Files But for Nerds…

TL;DR

The biggest AI company in the world just got its emails read out loud.
A new 10,000-word info dump, The OpenAI Files, reveals how the company went from nonprofit AGI guardian to a $300 billion closed-door operation shaped by investors, NDAs, and internal chaos. We’ve pulled out what matters.

A huge document leak just gave us the clearest view yet inside OpenAI. It covers everything from leadership behaviour and mission drift to secret deals and whistleblower concerns. The headline is simple: the original mission, to make AGI safe and available to all, has been quietly pushed aside in favour of scale, speed, and profitability. Because, why the fuck not?

The company is now a Public Benefit Corporation. On paper, that still sounds noble. But in practice, it means OpenAI is no longer bound to act in the public's best interest, and the board that was supposed to hold it accountable doesn’t have much power left. The “capped-profit” model is gone, the investor cap has been lifted, and a more typical Silicon Valley incentive engine is fully in place.

Three senior safety leads left in the past year: Mira Murati, Jan Leike, and Ilya Sutskever. Each issued public warnings about leadership and internal risk culture. Meanwhile, former employees report being locked into aggressive NDAs that threaten to strip their vested equity if they speak out. One employee, Leopold Aschenbrenner, was fired after raising the alarm about a serious 2023 security breach. That breach was never disclosed.

It’s not just about culture. Altman himself is at the centre of multiple questions, including whether he misled Congress about equity exposure, personally benefited from OpenAI partnerships, and quietly owned the company’s startup fund without disclosing it to the board. Combined, this paints a very different picture from the brand’s public messaging.

Other key details pulled from the docs:

• Altman listed himself as "Chairman of Y Combinator" in official filings despite never holding the title
  

• Former employees say OpenAI culture now prioritizes speed and secrecy over safety
  

• OpenAI lobbied behind the scenes to weaken the EU AI Act while publicly supporting regulation
  

• Whistleblower protections were reportedly waived as part of employment contracts
  

• Altman’s 7.5% stake in Reddit gained ~$50M after OpenAI inked a partnership with the company
  

• The company signed a $51M chip deal with Rain AI, where Altman is a personal investor
  

• The nonprofit board appears to have been sidelined in the company’s most recent restructure
  

• An internal review found Altman regularly gave different versions of events to different board members

Why this matters:

OpenAI is building the frontier of artificial intelligence. The tools it creates will shape how power, information, and influence move across the internet — and eventually, the real world. If transparency and safety are taking a backseat, we need to ask: Who is actually steering?

For me, I would rather Sam just come out and be a chad about all of this. I’d have way more respect for him and even trust him more if he said, yeah fuck it. I want to win, I want to be a gazillionaire, I hate Elon, and I will own 150% of OpenAI if I want to…

I can’t trust someone who makes out they have no material or financial interest in OpenAI succeeding whilst driving a McLaren F1…

Remember, folks, when in doubt, just be honest. Don’t piss on sombodies back and tell them it is raining…

Full files here: openaifiles.org

📈 Trending tools, models & apps this week

📋 LLM Leaderboard

📲 Trending tools & apps

🫵 Our Picks

What caught our eye this week…

• diabrowser – A designer-centric browser praised for its superb onboarding and workflow tailored to creatives.

• Gemini 2.5 Pro + NanoBrowser – Automates your browser tasks end-to-end using natural language via a free, open-source extension.

• Perplexity Labs – Replaces experts with AI to generate mood boards, McKinsey-style slides, landing pages and more from a single prompt.

• MiniMax-M1 – A powerful open-weight AI agent with 1M-token memory that combines “Manus, deep research, computer use” into a lovable assistant.

• Cursor for Slides – A next-gen presentation tool (“Cursor for Slides”) that outshines every other AI presentation assistant.

• Proactor – The first proactive AI teammate that listens, summarizes, researches, and acts in real time before you're even prompted.

🤓 Educational

Want to actually understand this stuff? Start here.

• Nerdy but Cool AI Tool – A slick, open-source AI assistant by Rohan Paul that geeks will love.

• How Anthropic Built Multi‑Agent Research Systems – A deep dive into Anthropic's new multi-agent research stack, showing how coordinated agent teams boost performance by over 90%.

• How Not to Lose Your Job to AI (80,000 Hours) – A practical guide by Benjamin Todd on high-value skills AI won’t replace and how to develop them to stay indispensable. You might even take your boss’s job after reading this.

• Innovative AI Widget – A stealthy, clever automation plug‑in by dcbruck that quietly enhances your workflow.
  

 🔥 Top Trending

Top trending apps this week that you have probably never heard of.

• Perplexity Tasks – A streamlined dashboard for managing your AI queries, research sessions, and saved tasks across Perplexity’s platform.

• Slashit – A versatile app that lets you slash through workflows with fast-access commands and AI-powered shortcuts.

• ComputerX – An all-in-one AI hub offering chat, search, and productivity tools in a unified interface.

• FetchAI (iOS) – A mobile app delivering instant AI access on your phone for on-the-go assistance and knowledge retrieval.

• HelloPetDreams – An AI-driven pet wellness platform offering personalized care plans and insights tailored to your furry friend.

👨‍🏫 Homework: Automate everything you can think of using n8n, no really…

I was nearly at the point where I'd given up with n8n. It's complex; there's a lot of nuance to it. However, the more you build, the more you research, and the more tutorials you watch, things start to click. Three weeks in, and everything has finally come to fruition, so you might want to persevere if you feel like giving up on it. From my self-torment, I thought I’d highlight the tools I’ve found useful and unlock some doors for us to help automate and speed up our data pipelines.

The first one is Web Search. LLM APIs don’t actively have this capability. On n8n a Perplexity tool is now available to be plugged into any agent you build, so you have access to live data. It's great because you get £5 free credits when signing up to test it, and the costs for Sonar are minimal for basic searches. It can be used for anything. I used it to search for recent press releases for raises and enrich company data. It has the full range of different models, so it can be applied to any agent you have, such as a deep research agent for your niche.

Another one I found useful this week is Firecrawl, which I imagine many people have seen. You can also tinker directly in the Firecrawl Playground to test its features firsthand. Firecrawl has several capabilities to explore, but here are the two I found useful:

• Scrape Capability: Scrape and crawl websites for data (website URLs, hyperlinks, page data, whatever you need from a page) into clear markdown. It is flexible enough to scrape any sort of web data that you want to retrieve.

• Extract capability: This allows you to extract semantic on-page data. I used it on several websites to collect project data that we would normally have to manually extract line by line.

You can then transpose or manipulate this data to create content, generate leads, or build complex workflows with multiple steps.

An additional tool for your arsenal is Apify. Apify is essentially a collection of purpose-built tools across loads of different “Actors” that have functions to extract information from all sorts of social media and beyond. You can search X users, Google trends, website data, LinkedIn job titles, YouTube metadata or even viral TikToks. Whatever you can think of for marketing, BD or research, you can scrape it using Apify, and again, you get free credits when signing up to test it.

I'm still using Claude to build my N8n workflow concepts; however, I don’t think this is necessarily optimised as much as it could be for selecting the correct nodes. More tutorials have come out for alternatives to help with builds. I particularly like Nate Herk’s YouTube channel; it’s been my go-to for learning new nodes and capabilities. He’s recently released a n8n workflow builder within n8n. I've just tested this week, and it's pretty cool. So it's an alternative approach if you don't want to pay for Claude individually and want to be on a pay-for-output work basis with all your costs under one roof through APIs.

I've also found that ChatGPT‑4o has been the goat for prompting structure and getting the correct JSON and JavaScript code to manipulate data within nodes on N8N. I’ve been using it as a sense check as I’ve been building to help conceptualise an intended result and found it so much more intuitive than Claude, so it's worth checking out.

🤝 In Partnership with OpenServ

Who’s Bridging the Agentic Framework Gap?

Agentic AI frameworks are a fragmented mess, forcing builders to juggle platforms and rebuild agents to keep up with daily tech drops. It’s exhausting.

OpenServ fixes this with an AI orchestration layer for seamless interoperability across all frameworks. Focus on what matters: building slick workflows and real outcomes, not platform-hopping.

No-code newbie or pro dev? OpenServ’s Playground Beta lets you spin up agents or tackle complex use cases with ease. Try it now at openserv.ai.

💸 Financials

• It’s that time of the weeks, folks – our latest instalment of AI taking our jobs. I know this is your favourite part of the newsletter. Well, for this week’s snippet, it is Amazon that are committing the dead, with Amazon CEO Andy Jassy revealing that he expects the company to reduce corporate jobs due to AI.

• Crosby has launched a new kind of AI-powered law firm, following a stealth $5.8million seed round led by Sequoia. While they continue to make AI software for lawyers, Ryan Daniels and Co are an actual law firm, that are using AI to provide legal service at break neck speed. And if you do happen to break your neck due to an accident at work, these guys will help, probably.

• Multiplier, founded by former Stripe business lead Noah Pepper, has raised a total of $27.5million in seed and Series A financing. The startup, now called Multiplier Holdings, aims to sell software to tax accountants.

• We have mentioned the rumours about ex-OpenAI CTO Mira Murati raising for her new company “Thinking Machines”, and I know what you are thinking… that is a cool name. Well, that was the name of this newsletter before she announced her silly little company before us. We’re not salty, she can take her recent $2bn raise at $10bn valuation and shove it up her old wahzoooooo.

• Sword Health, an AI-powered digital health startup, has raised $40m at a $4billion valuation, thanks to a funding round led by returning investor General Catalyst. It’s a whopping 33% jump from their $3bn price tag it earned a year ago.

• Base44, a six-month-old vibe-coding startup, has been sold to Wix for $80m it was announced this week. Israeli developer Maor Schlomo was the solo owner of Base44 (albeit he had eight employees working for him) and has made a pretty penny on his solo unicorn – with the deal being in cash, too. Note to editor Grant: We really need to start our company and make millions.

• Broadcom Inc. hit a major milestone this week, reaching a $1.2trillion market cap, making it a formidable contender among the elite technology stocks on Wall Street. It means Broadcom, a semiconductor manufacturing company, are now ranked as the seventh most valuable company in the S&P 500 Index, surpassing Telsa and Berkshire Hathaway.

• Those little fuckers over at Cluely have just gone and raised $15m from A16Z. Of course they have… Nobody even knows what they do yet but their Social Network announcement video was fire, so who cares…

🕵️‍♂️ 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.

👋 Until next week

Image generation has come a long way since 2022.

It was janky, had plenty of issues, but it was a start.

Now it’s advanced and pretty impressive… well, sort of. Someone on reddit showed how it’s come on leaps and bounds, even if Rick is missing half an arm.

Some other funny shit we found…

• FAANG is so 10 years ago: MANGO is the hot thing right now. Think it may be unfair Amazon lost their place but who really cares? and what the fuck were netflix doing in there?!

• Are you one of those bros demanding Cursor to fix their software? Well, maybe you should learn to code instead. Or maybe you should just ask Cursor to do it?

Asking Cursor to fix your code is just the beginning… take a look at the latest Y-Combinator batch… everything is claude-puter.

It’s good to be back to normal service. And don’t worry, I’ll be sticking round for good this time so you don’t have to put up with Matt… until next week when he covers for me again.

Oh and apparently we get Grok 3.5 very soon… let’s hope it doesn’t go on about white genocide again lawlllllll.

One final hit of encouragement from us… Keep doing what you need to do. Maybe one day you can be the next Jensen

Have a good weekend, boys and girls!

Sam, Grant, Mike and The Big Machines team.