God Help Us

Good morning,

Before we dive in I want you to know a big change is coming to this newsletter.
Over the last few years we have built an audience of more than 1000 “AI-aficionados”.
Most of you can’t get enough of our content.

A lot of work and research goes into this every single week.

And by now you’re probably thinking that we’re about to ask you for money.

But -plot twist- we’re doing the reverse. We’re going to give you even more value. To be honest I considered the money avenue just to be able to keep going.

Starting on Tuesday 29th of April 2025 we are launching "The Black Box”.

Every Tuesday, we will send one strange or powerful thing straight to your inbox.
Could be:

-A mind-bending AI tool
-A dangerous prompt or experiment
-A fringe paper or wild idea
-A GPT experience you didn’t see coming

It’ll be a very simple email drop but I promise you it’ll be interesting.

Back to the original content:

Breathe. Grab a coffee. We’re diving in.

Last week, two new models saw the light of day: OpenAI’s o3 and Google’s Gemini 2.5 Pro.

Two weeks ago, DeepMind dropped a research paper whispering the quiet part out loud: AGI — Artificial General Intelligence — won’t arrive with a lightning bolt. It’ll creep in, subtle and strange.

Tyler Cowen, the well-known economist, just announced that he thinks these two models have finally crossed “the line” and for him are “Artificial General Intelligence” .

Same message echoed in Ethan Mollick’s newsletter last week.
Mollick’s take was especially interesting — he claims our measuring tools are basically trash now. And that we actually don’t know anymore if we are in AGI territory or not.

Mollick introduces the concept of jagged AI — progress isn’t linear anymore. These models are now so sophisticated that we can’t even measure their sophistication.

We’ve built machines that compose symphonies of corporate jargon, vomit photographic landscapes, and casually spit out full-stack startup ideas over morning coffee — and yet -frustratingly - tthrow a basic riddle at them like Gollum did to Bilbo in The Hobbit and it might fail.

We are not only in “uncharted” territory but the map is “unchartable”.

Which explains why some people have the tendency to become skeptical of AI, even disillusioned. LLMs are unpredictable. One minute they're plotting a corporate strategy, the next they're flubbing a simple logic puzzle. And because of the fact they can’t solve the simple task - they’re being dismissed.

What we’re looking at isn’t a clean slope toward human-like intelligence — it’s a sawtooth path through bursts of superintelligence and baffling stupidity.

These new models — and whatever fresh beast OpenAI or Google lobs next over the walls of normalcy — don’t just solve problems. They orchestrate solutions. They wield tools, deconstruct goals, execute plans.

They’re agentic. Autonomous. Dangerously seductive. Like interns with God complexes and no impulse control. They can debug codebases blindfolded — but give them a logic puzzle and watch them sweat.

We’ve hit a point where these systems have become unmeasurable.
It’s like trying to measure a jet engine with a wind sock.

And so the definition of AGI — that shimmering idea of a machine that thinks like us but better — remains vaporous. Bloated with contradictions. Stretched thin by hype. Haunted by its own definitional ambiguity. And the financial woes of studios like OpenAI are not helping.

Ask five “experts” what AGI is and you’ll get a TED Talk, a manifesto, a shrug, two nervous breakdowns, and a YouTube documentary.

But back to Cowen. He argues the line was already crossed.

When a model can look at a photo and deduce where it was taken — then explain why in real time — or build a functioning economic model from scratch while narrating its logic like an economist with an espresso IV-drip, you have to ask: Not if we’ve crossed a line, but which line we crossed… without noticing.

And this is where things get psychedelic.

Because even as these systems strut through cognitive tasks like overconfident PhDs, they still demand constant babysitting. You can’t hand them the keys and walk away. They hallucinate. Fixate. Misfire. They are unpredictable in ways no regulatory framework is prepared to handle.

We are, quite literally, building the airplane while it’s already 30,000 feet in the air.

So what happens next?

Do we drift into a slow-roll future, with jagged systems nibbling their way into workflows and bureaucracies? (the preferred option)

Or does one of these agentic monstrosities crack the metacognitive code , becomes sentient and starts iterating on itself like a demonic startup founder with infinite seed funding? Add enough nodes to the neural network and consciousness might spark.

No one knows.

Not the CEOs. Not the red teamers. Not the authors of those 100-page safety papers filled with diagrams that look like LSD-fueled flowcharts.

AGI is like having Einstein in your basement - only he’s an expert at EVERYTHING- and you can use him for everything. Only, you are limited by your own intellect and imagination and you ask him to look up the best cat videos on YouTube.

That is what is happening.

But one thing is clear: those who learn to surf this jagged edge — to collaborate with these systems rather than stare at them in awe or run away from them in fear — will be the ones steering the ship.

That needs to be us.

Whether that ship coasts gently into the next economic boom… or crashes gloriously through every firewall we ever built, the wheel is already turning.

Welcome to the asymmetry.

Welcome to the noise.

Welcome to AGI.

Maybe.

Probably.

God help us.

AI News

• Google just launched Gemini 2.5 Flash, a fast and affordable hybrid reasoning model that rivals o4-mini and beats Claude 3.5 Sonnet on STEM tasks — now with a customizable 'thinking budget'. The new system lets developers control how much reasoning to use per task, optimizing for either cost or quality depending on the need.

• Profluent released ProGen3, a powerful AI that designs entirely new proteins — proving for the first time that scaling laws apply to biology like they do in language models. The model can create antibodies and gene editors more efficiently than traditional methods, potentially speeding up drug development and enabling new treatments.

• Meta’s FAIR lab open-sourced five new AI research tools focused on perception, 3D understanding, and collaboration — laying the groundwork for more advanced, real-world capable AI systems. The tools include SOTA visual encoders, a 3D spatial dataset, and a collaborative reasoning framework that boosts performance by letting AIs work together.

• Epoch co-founder Tamay Besiroglu launched Mechanize, a startup aiming to fully automate human work by training AI agents in simulated office environments. Backed by tech leaders like Jeff Dean, the startup focuses on replacing white-collar jobs — a move that’s already drawing criticism amid rising fears over AI’s impact on employment.

• Agentic coding tool Cursor faced backlash after its AI support bot made up a fake policy, sparking user cancellations and forcing the team to issue refunds. The hallucinated policy was blamed on a misunderstood security issue, highlighting the ongoing risks of relying on AI in customer service without clear oversight.

• DeepMind researchers proposed a new AI training method called “streams,” allowing AI to learn continuously from real-world feedback instead of just human data. The technique could push AI beyond imitation and into discovery, marking a potential shift from question-answering bots to systems capable of autonomous, open-ended learning.

• Anthropic analyzed 300,000 real conversations with Claude to map how the AI makes moral decisions, finding that its values shift depending on context. The study identified 3,307 unique values—like helpfulness, professionalism, and healthy boundaries—revealing that Claude’s ethics aren’t fixed, but adapt to different situations and user requests.

• The UAE announced it will become the first country to integrate AI into lawmaking, aiming to cut legislative drafting time by 70%. The new Regulatory Intelligence Office will use AI to suggest, review, and update laws, raising both optimism for efficiency and concern about bias and transparency in legal processes.

• DeepMind CEO and Nobel laureate Demis Hassabis told 60 Minutes that AGI could arrive in 5–10 years and cure all disease within a decade through AI-powered drug discovery. He also demoed DeepMind’s Project Astra assistant, which can recognize objects, read emotions, and run on smart glasses—hinting at how fast AI is moving toward everyday integration.

• Two undergrad founders from Korea just released Dia, an open-source voice model that outperforms industry leaders like ElevenLabs—despite having no funding. Built with help from Google’s TPU cloud, Dia delivers emotional tones, nonverbal cues, and expressive timing that rivals top commercial tools, showing how accessible advanced AI development has become.

• The Washington Post has partnered with OpenAI to provide article summaries and links directly in ChatGPT, joining over 20 major media outlets. The move highlights a growing divide between publishers embracing AI distribution and those, like the New York Times, still locked in legal battles over copyright and training data.

• Anthropic’s security chief says AI-powered “virtual employees” with login credentials and memory will join corporate networks within a year—raising new cybersecurity concerns. These AI workers will require account monitoring and access controls like human staff, signaling a shift in how companies must protect digital infrastructure in the age of autonomous agents.

• OpenAI just released its gpt-image-1 model to developers via API, powering ChatGPT’s viral image generator and opening it up for third-party tools like Adobe and Canva. The model allows for high-quality visuals, text rendering, and editing, with pricing based on tokens — enabling OpenAI to further embed itself across the creative tech ecosystem.

• Microsoft launched two new Copilot agents, Researcher and Analyst, aimed at transforming data analysis and research workflows — part of a bigger push toward AI-human hybrid “Frontier Firms.” According to Microsoft’s report, early adopters of AI are already seeing major productivity and morale gains, with agent-based work predicted to become the norm within five years.

• Over 30 AI experts and ex-OpenAI staffers urged state attorneys general to block OpenAI’s nonprofit-to-for-profit shift, saying it threatens the company’s original mission. The letter, backed by major names like Geoffrey Hinton, warns the restructuring could redirect AGI development from public benefit to profit — though the $40B SoftBank deal tied to the change is still moving forward.

Quickfire News

• OpenAI’s o3 model scored 136 (116 offline) on the Mensa Norway IQ test, marking the highest recorded result, surpassing Gemini 2.5 Pro.

• Chatbot Arena, UC Berkeley’s AI model evaluation platform, is becoming a standalone company under the name LMArena.

• Perplexity struck a deal with Motorola and is in talks with Samsung to make its AI search tool the default assistant or app on their smartphones.

• xAI’s Grok chatbot introduced memory support for past conversations and launched a Workspaces tab to help organize chats and files.

• Alibaba released Wan 2.1-FLF2V-14B, an open-source model that takes a first and last image frame as input to generate high-quality, coherent video outputs.

• Deezer reported 20,000+ AI-generated songs per day, using AI tech to help identify and filter synthetic audio content on the platform.

• OpenAI reportedly considered acquiring Anysphere, the maker of Cursor, before engaging in $3B acquisition talks with rival platform Windsurf.

• OpenAI’s new o3 and o4-mini models were found to hallucinate more often than older models, based on both internal and third-party testing.

• Google released a new version of Gemma 3 using Quantization-Aware Training, allowing the 27B model to run efficiently on consumer GPUs without losing performance.

• Sam Altman shared that OpenAI has spent “tens of millions of dollars” in compute processing user prompts with polite phrases like “please” and “thank you.”

• The Wikimedia Foundation partnered with Google’s Kaggle to release a dataset that helps AI developers avoid scraping Wikipedia, promoting proper use of its content.

• MIT researchers published a new method using sequential Monte Carlo for AI code generation, letting smaller models outperform larger ones by dropping weak outputs early.

• OpenAI added a Flex mode to its API pricing, letting developers cut costs in half for o3 and o4-mini calls in exchange for slower response times.

• Huawei is preparing shipments of its new 910C AI chip, which is designed to rival Nvidia’s H100, aiming to address demand in China amid U.S. export restrictions.

• Amazon Bedrock users are reporting service limitations with Anthropic’s models, leading some to bypass the issues by accessing Anthropic's API directly.

• Elon Musk is seeking over $25B in new funding for the xAI-X venture, a move that could value the combined company at up to $200B.

• ElevenLabs introduced Agent-to-Agent Transfers, a feature that enables seamless handoffs between AI agents for handling complex, multi-stage tasks.

• The Academy of Motion Picture Arts & Sciences announced AI use is allowed in films, clarifying it will not impact eligibility for Oscar nominations.

• Anthropic released a new best practices guide for its Claude Code platform, offering structured advice on building successful agentic coding workflows.

• OpenAI head of product Nick Turley stated in Google’s antitrust trial that the company would be interested in acquiring Google Chrome if it were ever required to be sold.

• Apple revised its marketing for Apple Intelligence, removing "available now" wording after the National Advertising Division raised concerns about misleading availability claims.

• Character AI launched AvatarFX, a tool that lets users generate long-form talking avatars using just one photo and a selected voice.

• IBM and the European Space Agency released TerraMind, an open-source AI system that leverages satellite data across nine modalities to monitor the climate in real time.

• Cohere CEO Aidan Gomez joined Rivian’s board of directors, with the goal of expanding AI integration into the automaker’s vehicles and manufacturing processes.

• Motorola introduced SVX, a new AI device that merges a body cam, speakers, and assistant tech to help emergency responders reduce response times.

• Perplexity launched its Perplexity Assistant app for iOS, enabling users to issue voice commands, browse the web, and perform agentic tasks directly from their phones.

• ByteDance’s Dreamina introduced Seedream 3.0, a new text-to-image model that now ranks second on Artificial Analysis’ Image Arena Leaderboard, just behind GPT-4o.

• OpenAI is projecting revenue of $125B in 2029 and $174B in 2030, with growth expected from AI agents, product expansion, and increased API and user adoption.

• NVIDIA released its NeMo microservices suite, designed to help enterprises build custom AI agents using company-specific data pipelines and flywheels.

• BMW will begin integrating DeepSeek’s AI models into its new vehicles in China later this year, becoming one of the first major automakers to adopt the startup’s technology.

• Tempus AI is collaborating with AstraZeneca and Pathos to create the largest multimodal foundation model focused on cancer treatment research and discovery.

Like newsletters?

Here are some newsletters our readers also enjoy:

Closing Thoughts

That’s it for us this week.

If you find any value from this newsletter, please pay it forward !

Thank you for being here !