What We Think

Good morning,

After last week’s article, I got an email from a reader to ask me where I stand on the issue of where AI development is going and my advice going forward.

I’m of course extremely enthusiastic about the technology on the one hand and on the other hand nobody likes to get lied to.

You can’t deny that the technological advancements are quite rapid. This week, Anthropic showed off their chatbot Claude’s new capability : it can now “use your computer”. It can fill in forms, look up websites , look at the source of a website, …

Once again, you can argue this is more of a party trick but also an impressive one.

Because every week there is a new advancement , you get a little desensitized to it. We’re frogs being boiled in water and I personally think that water is getting a little hot.

All things considered - this is my official stance :

AI will improve way faster than everyone thinks and change society way faster than everyone thinks.

We’ll get to that in a second but regarding my particular advice : even if we’re not in science fiction territory yet - you should start thinking about incorporating AI into your work today.

Therefore - I urge you to embrace Claude and GPT and develop the habit of having these tools open all the time. They hallucinate and make mistakes, yes, but if you know how to handle them you will work 10x faster than everyone around you and the quality of your work will generally be higher as well.

It does require a mindset shift because in the beginning - it feels like cheating. But working with it daily you will soon discover what it can and cannot do.

Some things I don’t let it do - for example the words you are reading now have not been AI-generated.

But for other things , like analyzing data, of course I use it.

Small disclaimer : I never use it a certain clients where potentially client data might get leaked. That is a big no-no - so do beware that you don’t do that.

But why do I think it will approve faster than everyone thinks ? I didn’t come to that conclusion myself but let myself be convinced by two individuals in particular. Almost eight years ago - my slight obsession with AI research started after watching a TED talk by Sam Harris.

Harris is a neuroscientist and at that time he was very prevalent in other fields - mostly as an outspoken atheist going on tour with the likes of Lawrence Krauss and Richard Dawkins.

In this talk Sam Harris delivered a warning against AI. In 2016. That’s kind of impressive. It’s quite interesting that Sam Harris comes at it from the neuroscientist angle. Consciousness and intelligence are to him nothing more than neurological connections. Put enough neuron cells together and you get consciousness - seems to be his main stance.

This week I rewatched the video. It’s oddly prescient. He even talks about a potential pandemic (which we had in 2020)

The Talk

Harris starts off by saying that humanity is facing an existential crisis like never before yet most people “seem unable to marshal an appropriate emotional response to the dangers that lie ahead”. AI dangers are not categorized as horror, but as science fiction and therefore most people find it “kind of cool”.

But in fact imagine one day humanity receives a message from an alien civilization telling us they’re on their way and they will arrive within 50 years so please prepare. The arrival of sentient AI, Harris argues, will have the same effect on humanity.

And chances are that AI will not be particularly friendly towards us. And it’s not likely armies of malicious robots will wreak havoc on human civilization like in Terminator or put us in a form of “VR” like in the Matrix. No, it will probably treat us like we humans treat ants.

“We don’t hate ,” he explains, “but whenever their presence seriously conflicts with one of our goals … we annihilate them without a qualm. The concern is that we will one day build machines that, whether they are conscious or not, could treat us with similar disregard.”

Superintelligent AI is “inevitable”

Sam Harris is a neuroscientist and his angle for looking at AI is very interesting.

Harris explains that one only needs to accept three basic assumptions to recognize the inevitability of superintelligent AI:

1. Intelligence is a product of information processing in physical systems.

2. We will continue to improve our intelligent machines.

3. We do not stand on the peak of intelligence or anywhere near it.

Humans have already created systems with narrow intelligence that exceeds human intelligence (such as computers). And since mere matter can give rise to general intelligence (as in the human brain), there is nothing, in principle, preventing advanced general intelligence in machines, which are also made of matter.

But Harris says the third assumption is “the crucial insight” that “makes our situation so precarious.” If machines surpass human intelligence and can improve themselves, they will be more capable than even the smartest humans—in unimaginable ways.

Even if a machine is no smarter than a team of researchers at MIT, “electronic circuits function about a million times faster than biochemical ones,” Harris explains. “So you set it running for a week, and it will perform 20,000 years of human-level intellectual work, week after week after week.”

Harris wonders, “how could we even understand, much less constrain, a mind making this sort of progress?”

Harris also worries that the power of superintelligent AI will be abused, furthering wealth inequality and increasing the risk of war. “This is a winner-take-all scenario,” he explains. Given the speed that these machines can process information, “to be six months ahead of the competition here is to be 500,000 years ahead, at a minimum.”

Harris warns that if his three basic assumptions are correct, “then we have to admit that we are in the process of building some sort of god. Now would be a good time to make sure it’s a god we can live with.”

A second person to influence my perspective is Mo Gawdat. He’s relatively obscure, doesn’t have a lot of followers but was the head of research at Google and is now making the rounds talking about the future of AI.

Don’t listen to me , listen to them. They are smart people. Unlike Sam Altman or Satya Nadella they are not after your money.

So I leave you with both these presentations. Enjoy and welcome to the Brief !

By the way, there is now a referral program for this newsletter. For every 3 people that you get to sign up, we will gift you a 15€ Amazon gift card.

Scroll to the bottom of the newsletter for details!

AI News

• Archetype AI has introduced 'Newton,' a foundational AI "Large Behavior Model" capable of learning complex physics directly from raw sensor data without human guidance. Newton accurately predicts the behavior of unfamiliar systems, outperforming specialized AI in tasks like power consumption forecasting. This breakthrough represents a shift toward AI models that can autonomously understand and adapt to real-world environments, potentially replacing narrow-focus systems.

• Google updated its AI note-taking assistant NotebookLM, allowing users to customize AI-generated audio summaries and adding features like Background Listening. A new business version is in the works, offering higher usage limits and collaboration tools. The viral success of Audio Overviews, which turns content into AI-hosted podcasts, is marking NotebookLM as a standout AI feature, signaling growing user enthusiasm for AI-generated media.

• Worldcoin, the identity verification startup founded by OpenAI CEO Sam Altman, has rebranded to ‘World’ and unveiled a new version of its iris-scanning Orb technology. The updated Orb is faster and easier to use, while the new World ID 3.0 protocol enhances privacy and includes features like deepfake detection. Despite its potential importance in verifying human identity against AI-generated content, the company faces challenges due to privacy concerns and regulatory pushback.

• Researchers at UCLA have developed SLIViT, an AI model capable of analyzing 3D medical scans with expert-level accuracy at 5,000 times the speed of human specialists. SLIViT can interpret MRIs, CT scans, and ultrasounds using just a few hundred training samples, thanks to its use of transfer learning. This breakthrough could significantly improve diagnostic speed and accessibility, particularly for providers with limited resources.

• Meta's FAIR division introduced several new research models and datasets, including Spirit LM, a multimodal language model that integrates speech and text, and an updated image segmentation tool, SAM 2.1. Other innovations include Layer Skip, which doubles LLM generation speed, and new tools for security testing and language model training. Meta’s commitment to open-source tools continues to shrink the gap between open and closed AI systems, pushing the boundaries of accessible AI innovation.

• Anthropic released new evaluations to test AI models' potential for sabotage, assessing risks like human decision manipulation, code sabotage, and undermining oversight. While current models like Claude 3 showed some capacity for sabotage, no immediate concerns were flagged. However, as AI continues to advance, the need for strong anti-sabotage safeguards will grow, signaling a proactive approach to preventing future risks.

• Microsoft announced new agentic capabilities for Copilot and Dynamics 365, allowing users to deploy or create autonomous agents to enhance business processes. Ten pre-built agents specializing in areas like sales and supply chain will be added to Dynamics 365, and users can build their own agents via Copilot Studio, which enters public preview next month. These agents, powered by OpenAI’s o1 models, aim to revolutionize workflows by automating tasks with minimal human oversight, marking a significant step in the rise of AI-driven business operations.

• Elon Musk's xAI launched the public beta of its API, allowing developers to integrate its Grok language models into applications. The API provides text generation, coding, and function-calling capabilities, with future support for vision models hinted. Priced competitively, xAI's API challenges established players like OpenAI and Anthropic, though its appeal may lie in offering less restrictive AI models, especially in combination with Grok's evolving features.

• AI startup Haiper released version 2.0 of its video generation platform, which allows users to create short 1080p videos, animate images, and repaint videos for free. While the platform offers impressive motion quality, it is currently limited by video length and quality filters. With backing from ex-Google DeepMind researchers and $19M in funding, Haiper adds to the growing field of AI video generation, signaling that AI-driven video content is poised for major advancements.

• Anthropic introduced a new "computer use" capability for its AI models, allowing Claude to navigate and control computer interfaces autonomously—typing, moving cursors, and executing commands across multiple applications. The Sonnet 3.5 upgrade enhances coding and tool use, outperforming competitors, while the Haiku 3.5 model offers similar high-end performance at lower costs. Though still imperfect, Claude’s ability to interact with computers signals a major advancement in automation, potentially reshaping the landscape for many startups focused on similar capabilities.

• AI startup Genmo released Mochi 1, the largest open-source video generation model, designed to compete with top closed models like Runway and Pika. Mochi generates 480p, 30fps videos with strong motion quality and prompt adherence, surpassing several top competitors in testing. With a higher-definition version, Mochi 1 HD, in the pipeline, and $28.4M in new funding, Genmo aims to push the boundaries of AI-driven video generation, signaling an intensifying race in the space.

• Ideogram launched Canvas, an AI-powered workspace that blends image generation and editing, featuring tools like Magic Fill for precise edits and Extend for expanding images while maintaining style consistency. This platform integrates seamlessly into creative workflows, allowing users to combine AI-generated and manual image editing. As AI continues to reshape design, Ideogram’s Canvas offers a powerful, user-friendly tool for both professionals and newcomers alike.

• Former OpenAI researcher Suchir Balaji has criticized the company for using copyrighted material in ChatGPT’s training, arguing that it violates copyright law and harms content creators. Balaji, who worked on GPT-4, contends that OpenAI’s reliance on internet data does not qualify as fair use. His comments come as OpenAI faces internal turnover, including the departure of senior policy researcher Miles Brundage, even as it hires new leadership, such as former White House economist Aaron "Ronnie" Chatterji.

• Google DeepMind has open-sourced SynthID, an advanced watermarking system for AI-generated content, already integrated into tools like Gemini and Vertex AI. SynthID embeds undetectable watermarks in text, audio, images, and video without affecting quality, helping distinguish AI-generated content from human-created work. By open-sourcing this technology, Google aims to set an industry standard for AI content transparency amid growing concerns over the authenticity of digital media.

• Runway introduced Act-One, a Gen-3 Alpha feature that allows users to map real human facial expressions onto AI-generated characters using just a smartphone video and reference image. This tool, which captures subtle expressions and movements, eliminates the need for specialized equipment and enables creators to apply performances across multiple characters. With this release, Runway continues to lower the barriers to high-quality animation, democratizing creative storytelling and visual effects.

Quickfire News

• The U.S. Treasury Department used AI to recover $1 billion in check fraud and prevent $4 billion in total fraud during the 2024 fiscal year, highlighting AI's growing role in combating financial crime.

• OpenAI expanded its partnership with Bain & Co. to develop and market AI tools tailored to various industries, with OpenAI reporting 1 million paying business customers.

• Meta is collaborating with Blumhouse and select filmmakers to test its Movie Gen AI video generation tools, gathering feedback before a planned public release in 2025.

• Researchers from Alibaba and Skywork introduced Meissonic, a small, open-source text-to-image model that outperforms larger models in generating high-quality outputs.

• Salesforce CEO Marc Benioff criticized Microsoft's AI efforts, particularly its Copilot assistant, calling it the "next Clippy" in an interview with Fast Company.

• OpenAI previewed its ChatGPT Windows app for paid users, which includes file and photo interactions, model enhancements, and a companion window mode.

• Perplexity is reportedly discussing a new fundraising round that would double its valuation to over $8 billion, according to the Wall Street Journal.

• Apple internally believes its AI technology is more than two years behind industry leaders, according to Bloomberg insider Mark Gurman.

• Midjourney will release a new tool this week allowing users to edit uploaded images using its AI model, alongside new retexturing features, with initial access limited to a smaller testing group.

• AI and quantum tech startup SandboxAQ is seeking new funding at a valuation of over $5 billion, with backing from former Google CEO Eric Schmidt and Salesforce CEO Marc Benioff.

• Publisher Penguin Random House updated its global copyright notices, now explicitly prohibiting the use of its texts for AI training, which will be included in all future titles.

• News Corp filed a lawsuit against Perplexity, accusing the AI startup of massive copyright infringement for reproducing content from The Wall Street Journal and New York Post without permission.

• IBM released Granite 3.0, a suite of open-source enterprise AI models that offer enhanced performance and improved safety features.

• GE HealthCare introduced CareIntellect for Oncology, an AI tool designed to consolidate patient data into a single view to help clinicians monitor cancer treatment progress.

• ByteDance, TikTok's parent company, fired an intern for allegedly sabotaging AI models by injecting malicious code, causing millions of dollars in damages.

• Alcon Entertainment sued Tesla and Warner Bros. Discovery, accusing them of using AI-generated images from Blade Runner 2049 in a Tesla Robotaxi promotion without authorization.

• Qualcomm unveiled the Snapdragon 8 Elite chip, featuring enhanced on-device AI capabilities that promise better performance and energy efficiency for future smartphones.

• Runway launched Act-One, a new feature that generates expressive character performances from a single video and image without the need for motion capture or rigging.

• Stability AI released Stable Diffusion 3.5, introducing Large and Large-Turbo models that enhance customization, efficiency, and output diversity.

• Cohere upgraded its Embed 3 model with multimodal capabilities, allowing enterprises to perform retrieval-augmented generation (RAG) searches across text and image content.

• Chipotle introduced a conversational AI hiring platform called "Ava Cado," claiming it can speed up the hiring process by up to 75%.

• Asana launched AI Studio, a no-code platform that allows teams to design and deploy AI agents for automating business workflows.

• Canva introduced Dream Lab, an image generator powered by Leonardo AI, alongside new AI features added to its Visual Suite.

• Inflection AI launched Agentic Workflows, enabling enterprise systems to take trusted actions for various business use cases.

• Apple rolled out Apple Intelligence features, including ChatGPT integration, Visual Intelligence, and creative AI tools like Genmoji and Image Playground, in its latest developer beta for iOS 18.2, with a public launch expected next week.

• OpenAI researchers introduced sCM, a model that achieves diffusion-level image quality in just two steps, reducing generation time by up to 50x.

• Qualcomm entered a multi-year partnership with Google to develop generative AI-powered in-vehicle technology, such as voice assistants and self-driving features, leveraging Snapdragon and Google Cloud.

• Over 10,500 artists, actors, and authors signed an open letter condemning tech companies for unauthorized use of creative works in AI training.

• xAI posted a job listing for work on autonomous agents, with engineer David Kim stating that “every human will have an AI agent capable of performing complex tasks on their behalf.”

Closing Thoughts

That’s it for us this week.

If you find any value from this newsletter. Please pay it forward and receive 5€ for every person you get to sign up in the form of an Amazon gift card. What’s keeping you ?