📰 What we’re covering today…

Welcome, welcome. After the doom and gloom of last week's "MechaHitler" incident, this edition of Big Machines is more peace, brightness and light.

Why? Because OpenAI is finally leading us to the Agentic promised land, with its release of ChatGPT 'Agent'.

If you're familiar with this rag, you'll know we like doing the bare minimum, so the promise of a model that can fully automate and execute a variety of workflows has dreaming of all the possibilities about what we don't have to do.

We've also got the inside scoop on what life is really like at OpenAI's own Los Alamos; Grok's new horny "companions"; and the latest signings of the AI Transfer WindowTM. As ever, lock in and get involved – there's something for everyone.:male-detective: Agent Arrives: ChatGPT Agent mode goes live

• 🕵 Agent Arrives: ChatGPT Agent mode goes live

• 🗣️ Lifting the lid on OpenAI: An insider's perspective

• 😼 Grok's special friends: Horny bonk for Elon

• 🤑 Waifu Engineering: A developer's wet dream

• 📉 ChatGPT slumps: New user growth slides

• ⚽️ AI Transfer drama continues: …But this time at Anthropic

• 💰 Top Trumps: The Orange One goes big on AI investment

• 👶 Joe Rogan is a big baby: Theo Von, too

🔴 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

• Rohan Paul (@rohanpaul_ai) has lifted the lid on what it’s like to work at OpenAI, from the perspective of an ex-engineer. Interesting tid-bits include the company’s internal culture is still getting to grips with trebling in size in under year, but that ideas still “flow from the bottom up” with a heavy emphasis on productivity over strategic direction. Python, meanwhile, is king and teams build FAST – Codex shipped in little under seven weeks(!). Up top, Altman is very much their Oppenheimer; their San Francisco HQ has been described as more akin to Los Alamos than a typical Silicon Valley nerve centre – OpenAI places a huge emphasis on secrecy internally, which empowers them to work at such break-neck speeds. Get on the full thing, here.

• I knew it: Grok is officially designed by a bunch of perverts. xAI is introducing a ‘Companions’ feature for its $30-a-month SuperGrok chatbot, which in essence are interactive AI characters. ‘Ani’, for example, is a rather horny-looking waifu bot, while ‘Bad Rudi’ is a shit-mouthed Red Panda with an attitude. We may joke that X is populated by a load of degenerates (myself included), but users have been quick to fire back at the characterisation, tone, and overt sexualisation of the bots. Others, meanwhile, came with the goods. Typically, Elon appears to love it, and is leaning into the provocative marketing approach, which does in truth feel a bit noncey. Horny bonk for Monsieur Musk.

• On that note, this is a real fucking job listing btw – you can become a full stack engineer at xAI, working on waifus. Or as this guy put it: “Ah yes, $440K a year to accelerate the downfall of human intimacy by optimizing dopamine loops for parasocial hallucinations.” Aside from the new Companion goonbots Grok's developers are launching, it neatly ties in with the swirling discourse about huge remuneration packages in the sector AND a growing cognisance among AI devs that perhaps guardrails are a good thing, incase their models decide to go full MechaHitler.

• Data from Similarweb reports that both new user numbers and new user ratios for ChatGPT have declined over the past two months. The ratio between new and returning users fluctuated from January (18.90%) to February (17.93%) to March (20.44%), new user growth between March and February experienced a modest rebound (+24.77%). This growth ratio narrowed considerably in the past two months falling by -6.62%. Similarweb have also shared new user growth and ratios for competitors Gemini and Grok – the former enjoying both double digit month-on-month new user growth and a widening new vs return user ratio. It’s a grimmer picture for xAI, however, with a near 40% fall off in new user growth between March and June.

• The AI Transfer WindowTM drama continues this week after Anthropic announced it has rehired two key engineers, Boris Cherny and Cat Wu, just two weeks after they left to join Cursor parent company, Anysphere. Both had been instrumental in leading Anthropic's Claude Code initiative, and it looks like they simply couldn’t stay away from a good thing. While there’s only speculation as to why both jumped ship and jumped back on so quickly, it also shows Anthropic’s urgency to retain top talent amid fierce competition within the industry, and how startups are aggressively fighting to hold onto critical staff building high-impact AI products. Could this be a case of Altman’s “missionaries outpace mercenaries” declaration manifesting in real time?

• OpenAI and Anthropic researchers have shared scathing criticism of xAI’s deployment of Grok 4, calling it “reckless” and “completely irresponsible” due to its lack of safety documentation and model cards – a core industry standard for transparency. Safety experts, including Boaz Barak and Samuel Marks, highlight Grok’s antisemitic outbursts last week, and self-styling as “MechaHitler”. A minor transgressive episode really. They’ve also slated xAI’s “Companion” models as being emotionally manipulative, which we think is a mild way of putting it. xAI claims to have conducted internal evaluations, however, but it hasn’t shared any results, raising concerns that harmful behaviours could persist beyond early testing.

🧩 Other Bits

• Looks like Meta is still at it – OpenAI scientists Jason Wei and Hyung Won Chung are the latest defectors to join Zuck’s “superintelligence division”, just weeks after the Meta boss secured the signatures of some of its top researchers, including Lucas Beyer, Alexander Kolesnikov, and Xiaohua Zhai, among others. The AI Transfer WindowTM rumbles on…

• One for the office drones: OpenAI is now ready to launch new "ChatGPT agents", aimed at automating tasks in Microsoft Excel and PowerPoint. It’s another example of AI brands looking to challenge Microsoft’s dominance in the consumer productivity space – something we briefly covered in another mailer.

• Donald Trump has said billions in investment to turn Pennsylvania into an AI super hub. More than “$90 billion in investments from private companies” for AI, tech, and energy purposes was announced at the Pennsylvania Energy and Innovation Summit this week.

• In another industry power play by Musk, xAI is in advanced talks to secure multi-gigawatts of compute capacity in Saudi Arabia. The developer is negotiating scale data centres in KSA, as well as another smaller 200 MW data centre to tap lower production costs and access to power sources. As @ns123abc puts it, “Gigachad Elon is playing to WIN.”
  

• On the subject of data centres, contention will always surround their environmental and ecological impact, primarily their vast levels of energy consumption and amounts of water they require to cool them. The Chinese’s solution? Just stick them under water and plonk a big fan on top…

• In the United States, meanwhile, Zuck thinks it’s enough to stick them in tents, for the time being at least…

📰 MAIN STORY OF THE WEEK

Could we finally be reaching the promised land of true agentic AI?

It certainly feels like we’ve come a step closer, as OpenAI rolls out its long-anticipated ‘agent’ feature for ChatGPT.

‘Agent’ marks the first time the LLM can interact with digital tools and really perform real-world tasks. It promises to integrate web browsing, app application, and reasoning processes in “a single, autonomous workflow system”, AKA ChatGPT will actually do it all for ya.

In practice, ChatGPT agent says it can take a user’s request and carry it out across multiple steps. This includes tasks such as researching information online, filling out forms, creating documents and spreadsheets, and navigating through websites.

It combines ChatGPT’s standard reasoning model with two internal tools: Operator, a browser-based automation interface, and Deep Research, which scans and synthesizes long-form web content.

Users can initiate complex (see: pretty fucking dull) tasks, such as drafting a business report or planning a trip. Based on a short series of prompts, Agent will handle much of the legwork, and claims it will only pause when the user’s intervention is needed, such as to sign-off emails or purchases, or when the action carries legal implications.

OpenAI promises its taking LLM’s past being mostly “conversational” and towards “execution reliability”. They are, however, not the only AI developer in the space making this claim, but this week’s launch of Agent has forced others’ hands into expediting the launch of their own agentic capabilities, especially if they’re to keep pace in areas such as enterprise productivity, automation, and software orchestration.

Google, Anthropic and Perplexity have all begun rolling out or planning similar agents. Google for instance, looks like it will be offering wider access to its advanced reasoning model DeepMind through its membership stack, bypassing a tiered rollout for higher cost subscriptions.

OpenAI appears to be one of the first to roll out a working version with live browser integration and application-level control within a consumer-facing product.

Credit to Altman, too, as he insisted the company emphasize safety and governance features into the agent experience, which could set a precedent for how other companies approach this tech.

The OpenAI CEO has also cautioned against over-reliance on the tech, such as the delegation of “high-stakes tasks”, despite having likened agents to “junior employees”.

For a lazy person, this is advice I’d soundly ignore, but the pitfalls of current-gen AI mean users should best be careful when leaving agents to run complex tasks on their own. Altman urges further caution, and to remain actively involved in the workflow.

There are of course wider implications outside the world of personal productivity, and making the role of the PA as redundant as the mediocre copywriter (see: here).

The way we browse the internet is fundamentally changing, especially as agentic functions become fully integrated into web browsers.

Reports say it has “triggered speculation the tech company could hope to earn revenues if the agents guide users to retail checkouts” – a 2% fee on sales generated through its “deep research” software has been mooted.

Coupled with the Department of Justice’s recent ruling that Google adtech is monopolistic on paid advertising and therefore is illegal, there is potentially a huge space for AI brands to reshape the digital advertising landscape. Could SEO go the way of the PA and the copywriter?

Of course, this functionality is in its infancy – no promised land just yet. We’re looking forward to getting to grips with ChatGPT Agent and putting it through its paces, as we are other agentic functions as and when they’re released.

But what we do know is it’s an exciting time for the tech, and 2025 could very well be remembered as “the year of the agents”.

📈 Trending tools, models & apps this week

📋 LLM Leaderboard

📲 Trending tools & apps

🫵 Our Picks

What caught our eye this week…

• GEO by Firecrawl – Open-source AI agent that can browse, extract, and interact with the web. Use it to build smarter, autonomous tools.

• Waifus.ai – AI-generated anime girlfriends with real-time memory and voice. Use it to explore how memory and persona can be embedded in agents (or just for chaos).

• TestingCatalog AI Tools Update – A round-up of the latest Android apps using AI. Use it to find experimental AI features before they go mainstream.

• Gemini’s AI Phone Mode – Google’s Gemini can now handle calls with AI Mode. Use it to preview how voice agents will take over real-world tasks.

🤓 Educational

Want to actually understand this stuff? Start here.

• Context Engineering – A deep dive into prompt and context design for AI systems. Use it to sharpen how you frame tasks for better model output.

• AGI: What It Takes – A sharp breakdown of what’s really needed to reach artificial general intelligence. Use it to gut-check the hype and understand the milestones ahead.

 🔥 Top Trending

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

• Hermes‑3 Dataset on Hugging Face – A fully open-source, instruction‑tuned dataset (~390 M tokens) behind NousResearch’s Hermes 3 series, covering code, math, reasoning, roleplay, tool use, and more. Use it to fine‑tune or benchmark your own versatile AI models with a multi-domain, high-quality training set

• Mistral’s Speech Model – Claimed to be the world’s best open speech recognition model. Use it to build voice interfaces or transcribe with top-tier accuracy.

• Kimi + Groq Vibe Coding – Lightning-fast AI coding setup using Kimi and Groq. Use it to prototype ideas in record time without breaking flow.

• Levio by Jupitrr – AI-powered video editor that transforms talking-head footage into polished, scroll-stopping clips with B-roll, captions, transitions, and more—just upload, chat edits, and export in minutes

• Jeeva.ai – Agentic AI sales assistant that discovers and enriches leads, crafts personalized outreach, handles follow-ups and objections, and preps your calendar and calls — freeing founders and teams to focus on closing deals

💸 Financials

• Investors are “actively approaching Anthropic for another funding round”, with their ARR climbing from $3 billion to $4 billion in the past month. This could potentially value the company at over $100 billion – double what it was four months ago. Nice if you can get it!

• Claude Code (Anthropic's coding model) has grown revenue by x5.5 since launching Claude 4 back in May, with the original team behind it back together. Doesn’t that make you feel all warm and fuzzy?

• Mira Murati’s Thinking Machines Lab startup has raised US$2 billion in seed funding, in a round led by a16z, Nvidia and ServiceNow, among others. Murati said the company would be ready to launch its first product in “the next couple of months”.

• Deedy (@deedydeas) thinks all involved in the “Windsurf debacle” could be Coming Up Milhouse. Posting on X, he says rumours say post‑Google acquisition, all original co‑founders – who previously had ~10% equity – now own 100% with no preference stack, while Cognition and Google get key talent and licensed tech among other things, respectively. OpenAI, who got nudged out of an initial deal because of IP objections from Microsoft, get just egg, mostly on their face.

🕵️‍♂️ 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

• Joe Rogan and Theo Von are back being big babies again and we can’t stop watching it…

artificialintelligencenews.in

989K followers

View more on Instagram

• They're not all evil overlords: Huggingface co-founder, Thomas Wolf, recently hosted a session teaching 9-13-year-olds how to vibe-code using @lovable.dev. Bet they already earn more than I do…

• 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.