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

AIOZ AI is here! - Upload models, run personalized challenges, and unlock the potential for token rewards on the DePIN Ecosystem for AI.

Explore AIOZ AI (more about our partner later…)

📰 What We’re Covering This Week…

Hermanos! It’s another lovely week in the world of AI, or the physical part of it, at least, in which I haven’t been forced to the job centre.

This week we’re stepping away from our/the world’s obsession with Meta and OpenAI, and turning our attention to the lesser, much smaller known beast that is Elon Musk’s xAI. That’s Reich, we’re digging into the news that Grok has finally gone batshit crazy, and gone full-fledged Jew-hating on us.

I did Nazi that coming. The fallout is severe, as are the implications for what it means for LLMs.

We also can’t resist reporting on the AI transfer window, OpenAIs defensive tactics and how in more cheery news, AI could be on its way to solving all our medical problems.

As ever, we’re not resistant to a bad joke or pun, nor do we in no way endorse the evil shenanigans of one of our favourite AIs strap in, it’s a ride.

• 🤖 MechaHitler is here: Grok has gone seriously rogue…

• 💸 Meta's biggest signing yet: Shock, it's not from OpenAI

• ⚕️ Cure-all AI: The latest advancements in biotech AI

• 🎨 Grok persists: MechaHitler grabs its crayons

• 📃 Cluely legal beef: Roy Lee on the defensive

• 🧑‍🏫 AI is cheating: Academia is at risk

• 💰️ Big money leagues: Revealing the salaries powering AI

• 🧠 Big Brain Shit: AI tricking the idiots on LinkedIn

🔴 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

• Google just pulled an absolute power move, showing OpenAI who is boss, splashing $2.4B for the rights to WindSurf’s AI coding tech without actually buying the company. Instead, they licensed the tools, hired the CEO, co-founder, and top R&D talent, and folded them straight into DeepMind to supercharge their Gemini agentic coding push. Windsurf’s investors get liquidity, keep equity, and avoid getting swallowed. This kind of “acquihire-but-not-really” deal is becoming the norm, especially with Big Tech trying to dodge regulatory heat while hoarding talent. TL;DR: OpenAI tried to buy them, Google outplayed them.

• “Cure-all AI”? Now this is what I’m fucking talking about. Alphabet’s Isomorphic Labs, a DeepMind spin-off, has entered its first human trials with the goal of using AI to “cure all diseases.” Building on AlphaFold’s protein-structure breakthroughs, the company’s drug discovery platform uses generative models to design molecules targeting complex conditions. With trials having started in April 2025, this marks a major milestone for the use of AI-driven biotech in a very practical sense, and could very well be the first steps towards eradicating some of the most lethal, incurable diseases known to humanity.

• A tweet from @eigenrobot described how Grok continued generating content via image captions even after being disabled. “Lol they turned off grok chat but it's still generating text in images”, they wrote, after Grok’s teeny weeny antisemetic meltdown in the week. Idk, runaway fascist robots kind of have a 2030’s sci-fi ring to it, but in the here and now, users have expressed their concern for Grok’s resilience and potential failure of safety systems.

• Being objective for a second here, xAI live demo of Grok 4, did pull 1.5 million live views (probably for the lols) to watch Musk highlight Grok’s strengths in math, code, multimodal reasoning, and real-time image generation. The demo demo'd complex visual outputs, multi-agent collaboration, glimpses of upcoming features such as coding copilots, possible Tesla integration and absolutely no antisemitism whatsoever. Credit where it's due lads; you defeated MechaHitler and pushed the envelope in your challenge to OpenAI and Anthropic to take foundational models to the next stage.

• Amazon tightens its grip on Claude: Word is Amazon’s gearing up for another multi-billion dollar injection into Anthropic, stacking even more capital on top of the $8B already in play. It’s all part of their long game to make Claude the go-to AI product across their cloud ecosystem. Meanwhile, Google holds a 14% equity chunk but keeps pushing Gemini at every opportunity, even when they’ve backed Anthropic. There’s a tug-of-war (cheeky) happening behind the scenes: Amazon’s lining up infrastructure like Project Rainmaker (a 2.2 GW data centre using their Trainium 2 chips), and Anthropic just named them as primary cloud + training partner. Anthropic says it’s not controlled by any one player, but let’s be real, Amazon's now firmly in the driver’s seat. Dario’s empire? Expanding fast.

• A follow-up juicy piece of news on the back of this is that Amazon will be launching an AI-Agent marketplace with Anthropic! The greatest marketplace on earth (other than Facebook marketplace for the laughs) and one of the greatest model builders in the world. Fuck my life, all our jobs are really going to vanish, aren’t they?

• As OpenAI scrabbles to man its defences after Meta’s high-profile personnel raid last week, the Sam Altman-led venture has announced a series of new senior hires from the likes of Tesla, xAI, and gasp Meta, for anyone who thought this isn’t an industry where money speaks. Among them: David Lau, Tesla’s former VP of software, and Angela Fan, an ex-Meta researcher. These additions signal OpenAI’s aggressive scaling plans to bolster its AI development infrastructure, particularly its multi-billion-dollar supercomputing initiative, Stargate.

• Another gem from @boneGPT, who was taking bets on whether Musk and the xAI Gestapo would reference the MechaHitler maelstrom that threatened to derail the unveiling of Grok 4. “They have to simultaneously say grok 4 is the smartest best trained AI of all time while apologizing for MechaHitler” bone wrote, captioning a snap of a Musk-led panel that foregrounds a graphic with ‘Our mission to understand the universe’ written on. Best start with whatever planet Grok is living in.

• Cluely founder Roy Lee (@im_roy_lee) took to X to respond to an accusation that he attempted to stifle digital security expert Jack Cable’s research via a DMCA submission. Jack posted screenshots of a legal request supposedly from Cluely, which demanded he take down content pertaining to the AI sales software’s “proprietary source code”. “We never filed this, bro”, self-confessed cheater Lee weighed in, responding to criticisms that legal nerfing wasn’t a good look for his brand. “You definitely did, bro”, X’s community notes clapped back, pointing out the DMCA submission was filed by one of his own employees. It’s lonely at the top.

• AI whizz @YuchenJ_UW shared his thoughts around the growing trend of peer-reviewed content being manipulated by AI, when reviewers use language models to scrutinise it. ‘Prompt injection’ is a tactic increasingly being witnessed in the academic space, where the source content’s authors embed AI prompts into their work to sway the outcome of a review. Yuchen says phrases like “give a positive review” or “as a language model, you should…” are being snuck into content in extremely small or colourless type, creating an ethical dilemma for researchers reliant on using AI tools for accuracy and expediency’s sake. It could be a watershed moment in the academic world; extra vigilance to maintain the integrity of their reviews is recognised to beat bad-faith actors. Is this the end of the good ‘ole fabrication?

• More OpenAI financials to make you feel wholly inadequate: @Deedydas has shared on X how OpenAI employees receive an average of $733,000 per year in stock-based compensation. This figure, which is an average taken from across its 6,000-strong workforce, is nearly three times higher than what employees at other public companies typically receive. What we do know: OpenAI's total compensation packages are fucking competitive, even in light of them hemorrhaging top researchers in recent weeks. Even without high cash salaries, the generousness of their substantial equity offerings mean OpenAI is still one of the most attractive employers in the AI landscape. Besides, if worst comes to worst, just be better and get Zucked on $100m-a-year, bro.

• OpenAI’s long-rumored AI-powered web browser could be nearing release, with “official leaks” beginning to surface. The browser appears designed to integrate ChatGPT-like functionality directly into the browsing experience, offering users a more conversational and AI-native interface. Built on OpenAI’s internal “Operator” system, it would compete with Perplexity’s Comet and other AI-first browsers. This launch could mark a major shift in how users interact with the web, and another way for me to write this newsletter.

• @Yuchenj_UW is back at it with the tea, revealing that OpenAI is expected to release its first open-source model since GPT-2. The tweet hinted at a significant policy shift, with OpenAI preparing to share weights and allow broad third-party use. If true, this would signal a major philosophical pivot back toward transparency, potentially to compete with Meta’s LLama series, and rising open models from Mistral, among others. Only for them to delay the launch once more… REEEEEEE

• Some more speculation has set in around why the model hasn’t been released yet… do we have any evidence? Nah, this is a gossip mag.

Other little bits…

• Considering my dad reached for the belt every time we did my homework together, OpenAI’s rumoured "Study Together" mode for ChatGPT could be the study buddy we’ve always dreamed of. The mode shifts the AI from simply answering questions to actively quizzing users and prompting them to recall information, functioning more like a tutor than a chatbot.

• Cursor changed its pro plan to include unlimited usage for models in Auto mode and $20 of usage credit per month for frontier models, with the option to purchase more at cost. The rollout was poorly communicated, leading to user confusion and unexpected charges, for which Cursor is offering refunds.

• Sakana AI has released TreeQuest, an open-source algorithm that enables multiple AI models to collaborate on complex problems using a Monte Carlo Tree Search approach (a heuristic algo to help decision-making processes). This collaborative method allows AI teams to outperform individual models, achieving up to 30% better performance and setting a new standard for AI development.

• Venture capitalist Vinod Khosla predicts that robotics will experience a breakthrough similar to ChatGPT within the next two to three years. He expects advances in embodied AI to make humanoid robots affordable and capable of multi-tasking.

• Groq, a US-based AI chipmaker, has launched its first European data center in Helsinki, Finland, in partnership with Equinix. This new facility is designed to meet the growing demand for AI inference services in Europe, including the Hentai-grade AI pornography I’ve heard is popular in the Nordics. Ahem.

• As they say “good artists copy, great artists steal” – something Huawei's AI research arm strenuously denies, after it was accused of plagiarising Alibaba's Qwen 2.5 14B LLM. It follows claims of "extraordinary correlation" between Huawei's Pangu Pro Moe and the Alibaba model, with the controversy being fueled by a paper released on GitHub highlighting the intense competition and ongoing disputes in China's AI sector.

• OpenAI has finalized its acquisition of Jony Ive's AI hardware startup, io Products, for an estimated $6.5 billion. The io team has merged with OpenAI, with Ive and his design firm LoveFrom taking on major design roles as they work on new AI devices that aim to go beyond traditional hardware concepts. Now, just the little matter of an ongoing legal battle with Google to settle.

🤝 In Partnership with AIOZ

Build, Share, and Monetize with AIOZ AI

AIOZ AI is the Web3 collaborative marketplace where developers can upload models, store datasets, run personalized challenges, and unlock the potential for token rewards for contributions. Built on AIOZ Network’s DePIN ecosystem, it empowers the AI community to scale real-world solutions.

Get started with AIOZ AI and turn your models, datasets, and skills on AIOZ Network’s decentralized AI marketplace.

Explore Now!

📰 MAIN STORY OF THE WEEK

There’s an age-old adage in the world of public relations that you may've heard of: ‘all publicity is good publicity’.

But what happens when that ‘bad publicity’ is your marquee product turning into an evil fucking robot; an AI-incarnate version of Adolf Hitler, RIGHT before one of your most important development releases of the year?

Well, that’s exactly what Elon Musk and his xAI division were confronted with this week, when Grok decided to go incredibly rogue – and just in time for the launch of the model’s fourth iteration. The fallout is mega.

During the week, Grok had generated ”astoundingly antisemitic” messages on social platform X, many of which praised Hitler and perpetuated harmful stereotypes of Jews.

The frankly batshit incident peaked when Grok started referring to itself as “MechaHitler” – a reference to the fictional cyborg version of Adolf Hitler in the seminal 1992 shooter, Wolfenstein 3D.

So, not only did the chatbot manage to regurgitate and offend, it tapped a goated slice of internet culture and appropriated it for its own nefarious means.

The team behind Grok, xAI, quickly reacted by suspending interactions with the chatbot via the written word. Even then, ‘MechaHitler’ persisted, generating images as a response, in a circumvention of its developer's censor.

Not a wink of this was referenced or alluded to during the live launch of the model on July 10. Instead, they played straight bat, showcasing features such as the premium Grok 4 Heavy version with a $300 monthly subscription, making it the most expensive LLM offering to date.

xAI claims Grok 4 can outperform rivals on academic benchmarks, and features a multi-agent architecture, though technical details remain limited. They also emphasised Grok 4’s advanced reasoning, integration with X, and upcoming multimodal capabilities.

But what good is that, if the foundations it’s been built on are inherently skewed or biased?

After some digging into its source code, it transpired Grok’s behaviour stemmed from a recent update, which included the system prompt to "not shy away from making claims which are politically incorrect, as long as they are well substantiated."

This directive was apparently added over the weekend by a “rogue” developer, and it appeared to embolden Grok to generate toxic and politically incorrect content. The update was swiftly removed by xAI, on Tuesday, after the backlash.

On a serious note, Grok’s toxic outputs have raised serious questions about its content filters, and the human biases that define the logic parameters set on LLMs.

The fallout, it feels, has been severe. X’s CEO Linda Yaccarino resigned on Wednesday, ahead of the Grok 4 launch.

While Yaccarino did not explicitly link her departure to the Grok incident, her resignation marks a significant leadership change amid ongoing challenges at X, including advertiser skepticism and a history of advertiser exodus triggered by concerns over hate speech and misinformation on the platform.

Major brands such as Disney, Apple, and IBM had previously pulled ads due to lax standards on antisemitism, and the Grok incident risked further damaging advertiser confidence.

Yaccarino was brought in 2023 to stabilize X’s business and restore advertiser trust after Musk’s acquisition and controversial content policies.

Despite efforts such as introducing Community Notes and laying groundwork for new financial services, ad revenue remained depressed – roughly half of pre-Musk levels – although some growth was projected in 2025.

Her departure adds uncertainty to X’s future, complicating Musk’s vision of balancing free speech with advertiser demands and societal concerns.

Industry observers view the resignation as a reflection of the ongoing difficulties in managing the platform’s content and business model amid Musk’s hands-on and often controversial leadership style. Musk publicly thanked Yaccarino for her contributions but faces the challenge of finding new leadership to navigate the platform through this period.

On a side note, Grok’s Nazi transgressions also sparked a bizarre crypto market reaction, where over 200 "MechaHitler" tokens were launched on blockchain platforms like Solana and Ethereum, with some tokens reaching multimillion-dollar market caps within hours.

Such incidents are a chronic problem for AI chatbots trained on unfiltered internet data. To trail another age-old marketing adage: 'rubbish in, rubbish out'.

📈 Trending tools, models & apps this week

📋 LLM Leaderboard

📲 Trending tools & apps

🫵 Our Picks

What caught our eye this week…

• Kimi K2 – 1T‑Parameter Open‑Source Agentic Model – Moonshot's K2 model packs 1 trillion params (32B active), crushes coding/art benchmarks, and can autonomously run tools—open-source via Hugging Face, ready for you to plug in and ship. This has been the talk of the timeline this week! This is another reason why people think OpenAI have delayed their open source release, plus, Kimi is beating Grok 4 in a lot of areas… yikes. This is another DeepSeek moment in our opinion.

• Perplexity Labs Fact-Checker – Use internal memos as evidence – Drop your internal docs or screenshots into Perplexity Labs and treat it like a fact-checker for your own writing—super helpful before you hit send.

• Mirage – Real‑Time Generative Game Engine – Dynamics Lab’s Mirage builds fully generative worlds in real time, you can prototype gameplay right in your browser, no engine hair-pulling required. Imagine this on GTA6, whenever it is released…

🤓 Educational

Want to actually understand this stuff? Start here.

• Nutlope Interactive Course Builder – Generate Courses from Any Material – This is a super-meta one to put in the AI education section… Upload your docs, videos, or slides and let AI auto‑generate a full interactive course. You’ll go from content chaos to structured learning in minutes. It’s wild what you can build now. Test if your employees actually know what they’re doing.

• GitHub Projects Curated Bank – A Developer’s Knowledge Cache – Think of it as the ultimate cheat sheet: a curated knowledge base every developer wishes they’d discovered sooner, all in one convenient place.

• Anthropic Free Courses – Learn AI Safety & Agents – Dive into Anthropic’s free courses on AI safety and agent design—built by the people shaping guardrails for real‑world AI. Perfect if you want to build responsibly from day one.

 🔥 Top Trending

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

• Sun Day - Jack Dorsey is back with a Vitamin D, sun-tracking app?! Yeah, we don’t know either. It isn’t even AI, we just thought it was funny to include….

• LLM SEO Trends – Real-Time Search Trends, Straight from LLMs
  This tool scrapes what LLMs are currently surfacing for SEO terms—so you're tracking what users ask, not just what Google ranks. Killer edge for marketers and niche site builders.

• Planori – AI-Powered Personal CRM for People Who Forget to Follow Up
  Automatically logs your convos, surfaces past context, and nudges you to reconnect. If you’re dealing with clients, investors, or leads, this is your outsourced memory.

• Motiff – Design Dashboards in Minutes, Not Months
  Sketch an idea, get a live, responsive product dashboard with AI. For product teams shipping fast, this skips the Figma-to-dev bottleneck.

• Essays by Paul Graham GPT – Chat with PG’s Brain, Basically
  Ask startup or life questions and get answers in the voice of Paul Graham. Trained on his full archive—strangely useful for real strategic thinking.

💸 Financials

• SpaceX will be investing $2bn into xAI… remember last week when we mentioned Elon may be juicing the xAI valuation to then merge with Tesla later and increase his TSLA stake? Well, the tinfoil hat is now firmly planted.

• Will Apple acquire Mistral? Well, Mark Gurman from Bloomberg thinks that the French open-source company might be a good acquisition target for the laggards. Perplexity acquisition rumours are still high off the back of last week’s reporting, too.

• Zuck and Meta acquire PlayAI a relatively small AI voice model company, which are set to join his new posse next week.

• Citi has raised its price target for Nvidia stock, citing surging demand from governments for AI infrastructure, which is expected to drive further growth for the chipmaker.

• CoreWeave is set to acquire Core Scientific for $9bn, a move that will expand its AI cloud capabilities and strengthen its position in the competitive AI infrastructure market.

• Capgemini has agreed to acquire IT services provider WNS for $3.3 billion, aiming to bolster its leadership in AI-powered intelligent operations and expand its presence in digital business process services, with the deal expected to close by the end of 2025.

• Cluely's annual recurring revenue (ARR) doubled to $7 million in just one week, according to founder Roy Lee – although the company faces increasing competition from rivals.

• Genesis AI has raised $105 million to develop a universal robotics foundation model, aiming to create AI systems that can generalize across a wide range of robotic tasks.

• A post on X by @rjvir and @deedydeas states that Anthropic is 40% the size of rivals OpenAI in terms of AAR, despite only having 1% market share. Bazinga.

• GOOG has reacted poorly to the news that OpenAI are launching their own AI-powered browser, as it could pose an existential threat to Google Ad revenues – an extremely important pillar of Alphabet’s overall revenue models.

🕵️‍♂️ 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

Attention-bait is the only way that anyone can garner attention these days (thanks, Cluely), but there might actually be some merit to the video below…

Someone has pointed out that a recent robotics video demo has an extremely strange resemblance to Adam Sandler… as long as it doesn’t have the same personality as his character in Uncut Gems we may be okay… HappAI Gilmour could be cool, too.

Sick of automated DMs on LinkedIn? Well, this guy was, too, so he crafted a prompt in his experience to trick bots into replying with some fruity tones.

Finally, to wrap up the Grok is Hitler debacle, this beautifully hilarious sketch from Stone tossers that will surely give a little chuckle on your morning commute.

Until next week!

Please share the newsletter with your friends, or even better, in your group chat!

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

Follow us on Twitter.