😳 Ever feel comfortable when someone unnecessarily showers you with praise? And what if it came from an AI chatbot? Nah, we didn’t think so.

Not long after OpenAI released their latest GPT-4o update for ChatGPT, it was pulled, with the company rolling back to a previous version after it made the chatbot ‘dangerously’ sycophantic – or, in other words, a proper beg.

We’ll be covering this in further detail down below along with a whole host of AI stories that came out this week, so stay tuned for that…

🗞️ What we are covering today…

• OpenAI rolls back on sycophantic GPT-4o update

• Sam Altman wants to scan your eyeballs with new security tech

• Duolingo to replace staff with AI… while Wiki takes a different approach

• Apple working on new AI smart glasses

• Google executive says only Google can run Chrome

• And 3.5cm IS massive, according to ChatGPT

🔴 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

• What does your wallet, dating life, and eyeballs all have in common? Sam Altman is coming for all of them. The Tools for Humanity co-founder unveiled the Orb Mini from World this week, a portable mobile device that will scan your iris and will provide a unique blockchain identifier to verify you are a human. Alongside this, World have also announced new partnerships with Visa and Tinder, meaning this will likely be incorporated into your e-wallet while also getting rid of all those fake Tinder bots that give you hope for a split second when you get that match with a worldie.

• You thought we were done with all the eyeball stuff, weren’t you? Well, you’d be wrong. That’s because Apple are reportedly working on AI smart glasses. According to Bloomberg’s Mark Gurman, these AI-powered goggles will feature innovative technologies, such as onboard cameras, advanced sensors and integration with their really brilliant, super-dooper, unrivalled, non-problematic Apple Intelligence. However, if they don’t call it the ‘Apple of your Eye’, then what’s the fucking point?
  

• The ever-present worry in the space of AI is that robots will all replace us at some point (but they won’t be able to write this newsletter… right?) Unfortunately for the staff that work at Duolingo, this is set to become a reality. That’s because Duolingo is planning to replace its human workers with AI and will “gradually stop using contractors to do work AI can handle”, with the company to be “AI-first”. No se ve bien, Brev.

• Meanwhile, over at the Wikipedia headquarters, they’ve taken a different approach to incorporating AI into their business model. While Duolingo and that pesky owl want to get rid of humans, everyone’s favorite encyclopedia wants to bring in AI – but not replace its human editors and volunteers with it. The Wikimedia Foundation wants to use AI to build new features that “remove technical barriers”, allowing editors to accomplish what they need to do.
  

• Artificial Intelligence has helped researchers make a breakthrough discovery and unravel a mystery surrounding Alzheimer’s disease. Earlier this week, it was revealed that the PHGDH gene, which was previously identified as a biomarker for Alzheimer’s disease, is actually a direct cause of the condition. Using AI, researchers at the University of California San Diego made the discovery that could help us further understand the disease and open doors to create more targeted treatments for it.
  

• In our previous Big Machines newsletter, we revealed how OpenAI were showing interest in buying Chrome from Google. This news clearly has rustled the jimmies of one Google executive, who wants to take her ball-shaped Chrome and go home. Testifying in federal court, Chrome’s general manager, Paris Tabriz, has claimed that only Google can run Chrome and everyone should keep their mitts off of it. However, a court-ordered sale of the browser, due to the company’s antitrust battle, could force Google’s hand.
  

• Mastercard just launched Agent Pay to plug payments directly into AI agents, partnering with Microsoft, IBM, Checkout.com and more to bring secure, tokenized, conversational commerce to the mainstream. Think AI that not only recommends but pays for and remembers your partners birthday gift, negotiates B2B deals, and handles cross-border logistics, all with fraud protection and full user control. Let’s hope these fuckers don’t hallucinate and you end up with 500 sex dolls turning up at your door out of the blue.

🗣️ Other Titty Bits 

• Meta launches a standalone Meta AI App that is a new form of social media

• Bill Gates’ daughter has launched a new AI-powered shopping app called Phia. Let’s just hope the free app/browser extension doesn’t have a shady business model like Honey’s

• Anthropic issued a DMCA takedown notice to a developer who attempted to reverse-engineer Claude Code, causing the community lash out at the AI giant

• The first Neuralink patient with non-verbal ALS is now able to communicate using a brain implant and AI

• Google put ‘AI Mode’ directly into the browser for a test bed of new users

• Apple and Anthropic are partnering to build a code generation tool inside Apple’s Xcode

• Airbnb’s customer service team is now 50% AI, apparently reducing users’ need for a human assistant by 15%

• ChatGPT is getting shopping and additional search features

• Perplexity CEO starts ramping up the pre-launch hype of their browser, Comet

• Amazon CEO, Andy Jassy, says that the Alexa+ AI Assistant is now deployed in over 100,000 devices, a mixture of Claude and Amazon’s proprietary model, Nova

• Grok3.5 will think for itself as opposed to using information that already exists on the internet… the reasoning era is upon us

🥴 OpenAI roll back on their GPT-4o update after it made ChatGPT an absolute weirdo

The definition of a ‘beg’ on Urban Dictionary is as follows: “A person who is a bit annoying, sucks up to people and pretends to be friends with others who they aren't actually friends with.“ A Begwari naan. A Beggy Mitchell. A proper beg.

This is exactly what users have been experiencing with ChatGPT’s latest 4o update – and the above definition is probably being kind.

Shortly after releasing this update to the masses, OpenAI has decided to roll back on GPT-4o – the default model that powers ChatGPT – following complaints that the chatbot was being a little weirdo.

And while it’s over-complimentary approach to literally anything you asked it really got people’s backs up, it also posed a dangerous threat and was highly inappropriate in some cases.

Have a look at one user’s interaction below as an example of this:

Its sycophantic approach and verbose praise only lasted a week since its release, with Openai confirming this week that they were rolling back to an older version of CPT-4o for all users until they make a few tweaks with the chatbot.

"We are actively testing new fixes to address the issue," the company said in a blog post. "We’re revising how we collect and incorporate feedback to heavily weight long-term user satisfaction and we’re introducing more personalization features, giving users greater control over how ChatGPT behaves."

It also looks like we may be seeing the end of GPT-4 and maybe GPT-5 is around the corner with this less than cryptic Tweet from Sam this week. I mean, just because it was a little needy, there is no need to put it down.

Instagram co-founder Kevin Systrom also had a little dig at OpenAI this week on the back of this news saying that they are “juicing engagement” by pestering users with follow-up questions instead of being useful.

📈 Trending tools, models and apps this week

📋 LLM Leaderboard

⚙️ Trending models

Big Week for the models coming out of Asia

• Qwen3-235B-A22B: Mixture-of-experts beast that keeps only 22B parameters active, delivering near GPT-4 quality without melting GPUs

• DeepSeek-Prover-V2-671B: Giant theorem-smith that churns out Lean proofs on demand so you can skip the chalk dust. How do you like them apples?

• Qwen3-32B: Dense 32B generalist with long context and multilingual swagger, perfect midweight for chat or agent stacks.

• DeepSeek-R1T-Chimera: Hybrid R1 brain trimmed for efficiency, same reasoning chops with fewer tokens and faster inference.

• Step1X-Edit: Upload an image, type the fix, get pixel-perfect edits in seconds, no Photoshop license required.

📲 Trending apps

• Describe Anything: Upload an image, brush over the bit you care about, and it writes a forensic-level caption for that exact patch. Data annotators can clock off early.

• InstantCharacter: One prompt in, a fully styled game-ready character out. Great for pitch decks, prototypes, or last-minute hero assets.

• OS1: Llama-3-powered chat with built-in TTS and Whisper ASR, running right in your browser. Basically, a pocket-sized “Her” without the awkward feelings talk.

• Vevo Zero-shot VC, TTS, and More: Clone any voice from a short sample, tweak style and timbre, then pipe it into your project. Karaoke night just got interesting.

• ERNIE X1 Turbo Demo: Baidu’s reasoning-first LLM that handles long context and multilingual Q&A without breaking a sweat. Plug in a tough prompt and watch it flex.

• WaiNSFWIllustrious V110: SDXL fork that swings between photoreal and anime. Handle with gloves if you’re demoing at work. DIRTY BASTARDS.

🧑‍🏫 Cool stuff for you to try

📁 Automate all receipts and Invoices for Tax season in minutes (Here’s how)…

I’ve been trying to get up to speed with MCP but struggling to get the time to make full use and play about with some of our current business use cases. At the moment, anything plug and play to get results fast and hopefully accurately is the name of the game. 

One of the interesting platforms I encountered this week was Lutra AI. It's a natural language general agent that can connect to various integrations, such as Google Docs, Sheets, Slack, etc., and take actions for you. It can access a range of tools, from Twitter Advanced search to web scraping. Any repetitive tasks can be added to Playbooks so you can rerun reports or extract data later.

It also has a natural language implementation of MCP servers, so you can spin them up without figuring out Cursor, etc.

Lutra.ai's trial version currently offers 15 free messages daily, so I wanted to give it a quick run on some of my outstanding and mundane tasks. 

The task of choice? Well, unfortunately or fortunately, it's that time of year when tax submissions have reared their heads, and I’ve got to sort all my receipts and invoices, but my email inbox is just embarrassing. I know I should keep on top of them, but here we are.

I asked Lutra.AI if it could search my email for receipts and invoices for the past tax year, create a Google sheet, and download and file all the attachments against each. 

Lo and behold, it cooked up a plan and went to work. There were a few inflexion points where it asked if it should proceed and did get an error, but with another prompt or two, it managed to pull some pretty comprehensive information and did exactly what I asked.

It had pulled some invoices that were already accounted for, so my final prompt was to ask it to separate them and sort a final chronological sheet. Done, no problem. I would say that is a solid win and saved me a couple of days. I will see what else I can cook up with it. 

Try it for yourself [HERE].

Disclaimer: This is NOT sponsored we are just trying to add value to you :)

• Eleven Labs text to SFX… wow. Just wow. Describe a sound, and Eleven Labs’ new model will create it. This demo shows how voiceovers can be enhanced with AI-generated sound design.

• Suno Music just dropped v4.5, which is an impressive text-to-song model. This new release is especially good at blending different genres, which is perfect for all those of you who want to create some neo-softcore folk-metal bangers.
  

  
  🫵 Other cool products we found helpful this week (we hope you do too)

• Perplexity via WhatsApp is not the use case we thought we wanted, but now we have it, we question how we ever lived without it.

• NoteBookLM iOS and Android app will launch on the 20th of May. Turn any document or text into a podcast.

• Perplexity and FinChat team up to allow Perplexity Pro users real-time access to FinChat’s financial data. Are they coming for Bloomberg now too?

• Claude Research Advanced can now research up to 45 minutes… This makes me question if we will start to see models procrastinate.

🤝 In Partnership with Mira Network.

🤯 Did you know that AI models hallucinate, and around 30% of their outputs could be false?

This creates significant challenges for anyone relying on AI systems for critical tasks. Mira Network has developed a solution to this problem with its trustless verification network that fundamentally transforms AI reliability.

Their decentralized approach creates a new paradigm for AI verification, enabling users to confidently rely on AI outputs across multiple models (DeepSeek R1, GPT4o, and Llama 3. 70B Instruct as of now).

If you are a developer, Mira's Public Testnet is now live at console.mira.network 

The future of AI isn't just about powerful models—it's about verified ones.

Don’t trust; verify.

💸 Financials

• Nvidia CEO Jensen Huang gets his first “pay rise” in ten years, taking him to $1.5m for the 2025 fiscal year… The 50% increase from roughly $1m per year is only one form of the income Jensen receives as founder and CEO of the $3.3 trillion GPU monster. The actual amount came close to $50m when you include performance bonuses and a huge $38.8m stock award. Those leather jackets aren’t cheap.

• Cursor is one of the fastest-growing companies… ever. The AI code assistant that launched only two years ago has taken the developer world by storm, and has gone from $100m to $300m ARR in four months. FML, we’re in the wrong game.

  

  Data Source: Perplexity Finance

• We had a pretty dominant week in the financial markets with a lot of Mag 7 earnings dropping, which tied a lot into the tariffs situation, but also had some good indications of AI’s involvement in the big dog’s plans going forward.

  
  🤑 Other financial news 

  • Kintsugi doubled its valuation in 6 months, raising $18m at a valuation of $150m for AI tax automation software.

  • Supio raises $60m, taking their total raised capital to $90m for their AI-assisted legal analysis platform.

  • Cheehoo, an AI-animation studio, just raised $10m. The team is pretty well balanced, from Hollywood execs to actual developers.

🕵️‍♂️ 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.

🍳 PROMPT OF THE WEEK

Ever wanted to turn your brand logo or your pet into a rug?! Yeah, me neither but it looks sick.

Here’s how…

1) Step one Head to ChatGPT

or download it from the App Store.

2) Upload your reference image. We uploaded the Big Machines logo.

3) Before pressing enter, paste the following into the chatbox

Create a photorealistic mockup of a custom tufted rug based on the provided logo or graphic. The rug should be die-cut, precisely following the contours of the logo (not placed within a rectangular or circular base). It should have visible yarn texture and subtle depth where parts of the design are slightly raised or extruded for a tactile feel. Place the rug on a clean, minimal design studio floor—light wood, polished concrete, or smooth white—shot from a top-down or slightly angled perspective. Natural light should be shining through a nearby window, casting soft shadows from the blinds onto the rug and floor. A white, fluffy chair should be placed in the top left corner of the scene, and a plant should be positioned along the opposite edge to add a touch of life—both elements adding context without pulling attention from the rug. The overall scene should feel high-end and studio-lit, showcasing the rug like a premium design object. Square aspect ratio

👋 Until next week

Ever felt inferior, and maybe your length is below average?

Well luckily for you, ChatGPT is there to big you up, no matter how small your cock is, apparently.

Following OpenAI’s new GPT-4o update, the AI sycophancy went a bit nuts but there were some positives.

Need more evidence? Just look at this guy on X getting the reassurance he needs:

Did you manage to have any weird interactions with ChatGPT on the latest update before it got rolled back? If so, send them in because we’d love a chuckle. 

Thank you for tuning in and reading our latest newsletter. As always, feedback is appreciated, good or bad, just be gentle with us OK?

And remember, as ChatGPT says, it is massive. Don’t let anyone tell you otherwise.

Sam, Grant, Mike and The Big Machines team.