Is Sam A Snake ?

Good morning,

The field of AI is heating up again. We are officially making the move from “chatbots” like Claude and ChatGPT to "agents.”. Microsoft, OpenAI, Google, and Anthropic have all announced new releases coming early in 2025.

Agents can be given a complex task and will go off and do it and present the result. You can ask an agent to write a thesis on a topic, and it will go do the research, write the paper, and neatly format it into a document. That’s the promise.

The step after this is where agents will come up with novel ideas and new science.
At least according to OpenAI’s Sam Altman.

But let’s look a little closer at Sam today.

This week I read “Supremacy - AI , ChatGPT and the Race that Will Change the World” by Parmy Olson. 

It’s a book about the 2 companies at the forefront of the AI evolution: Google Deepmind and the inevitable OpenAI. How they came to be, how they have evolved, and where they seem to be going.

It’s also about the founders of these companies: Demis Hassabis (DeepMind) and Sam Altman (OpenAI).

About their characters, flaws, and strengths; their origins and ambitions.

What really stands out here is that Hassabis is all about the science. A programming prodigy—creating commercial video games at the tender age of 15. He views life in terms of games and is also developing artificial intelligence through training them on games. DeepMind notoriously beat the world’s leading Go player and obliterated the best chess mind in the world a few years earlier (Hassabis himself is one of the 100 most successful chess players in the world). Hassabis also won the Nobel Prize in Physics for his work in AI together with Geoffrey Hinton a while ago. Hassabis is a true AI scientist.

As an aside, I also felt very connected to Hassabis, as he’s exactly my age and started his career on the same computer as I did.

Behold, the mighty ZX Spectrum—both Hassabis and yours truly started on:

I still remember vividly buying a book on “Basic” and trying to program games. Hassabis did the exact same thing. (Only he made games; I played them.).

But Hassabis was always in the shadows.
You probably don’t even know what Hassabis looks like. Let me fix that.

Altman is a different type of person.

He is more of a typical entrepreneur. He has the skill of bringing the right people together and is an incredible networker. He knows everyone in Silicon Valley—he goes to all the cocktail parties and is more of an evangelist of AI. He’s all talk and less content.

You know the type; we’ve all met them. These salesy types that try to look for the next conversation partner at a cocktail party as you start asking some probing questions.

He’s also everywhere. I cannot open up my YouTube feed without Altman staring back at me from another video at another conference. Hassabis is more reclusive.

The conclusion that I came away with is that Altman is actually pretty dangerous and clearly in it for the money. Nothing wrong with that unless you claim otherwise. He recently secured himself a nice 10-billion-dollar rewards package within all the outside investment. Even though he proclaimed in the US Senate he does this because he loves it—I have now 10 billion reasons to assume that is a lie. All his talk about the race to AGI is about getting investment for the race to AGI.

The success of OpenAI is also hugely due to its founder dangling the AGI carrot in front of investors.

About 6 months ago, I talked in this newsletter about the fact that 2025 is going to be the year of AI disillusionment because nothing groundbreaking will come out of the AI labs.

This week there was a lot of talk about the scaling laws being broken. AI companies are starting to discover that what they assumed about AI scaling is not true: that if you give an AI model more data, it will become more intelligent. That is not true. It will get smarter—not more intelligent. Big difference.

And you cannot get to AGI by going bigger.

There are not enough graphical cards in the world to reach for “AGI"—we're not getting there by increasing in scale. Both Hassabis and Altman have made statements along those lines during the past week. “We’ll have to get there through working on the engineering of the models, not the scale.”

This was confirmed by Dario Amodei - the CEO of Anthropic on the Lex Fridman show ( A whopping 6 hour deep dive on AI). Where Amodei basically admits the same.
But when pushed for the “when AGI?” question, Amodei (who looks like the love child of Hassabis and Altman) sees his company getting there in 2026 or 2027.

Watch him answer the question in the video below

Of course Altman—one upped him and made headlines this week by declaring he expects AGI to arrive in 2025. That was actually the headline in the news. BUT - th’s not exactly what he said. And this is indicative of what’s wrong with AI news and what is wrong with Altman. He makes these kind of opaque statements.

Interviewer: what are you excited about in 2025.
Altman: AGI. And having a kid.

It’s so typical for Altman to give these kinds of vague answers, knowing full well how they can be interpreted. Judge for yourself (it’s at 46:21)

“We basically know what to do to achieve AGI. The roadmap is clear …”

So yes - I’m conflicted and confused.

I see Altman for what he is.

I also desperately want it all to be true. Humanity needs that little spark of hope.
We’re not doing that great if you look around.

And either Altman and AGI is the savior of humanity or he’s the vilest sort of grifter you can imagine, abusing our hopes for a better world for personal gain.

Time will tell.

Welcome to the Blacklynx 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

• OpenAI plans to launch Operator in January, an agentic AI tool designed to complete multi-step tasks like booking flights or writing code with minimal oversight. Positioned as a major step forward from traditional chatbots, Operator will compete with similar tools from Anthropic, Microsoft, and Google.

• Researchers at Johns Hopkins University have trained a surgical robot to perform complex procedures by watching videos of human surgeons, achieving human-level skill in tasks like suturing and tissue manipulation. Using imitation learning and ChatGPT-style architecture, the robot learned through surgical footage and showed unexpected adaptability, such as recovering dropped needles without explicit programming.

• Humanoid robot artist Ai-Da made history at Sotheby’s Auction House, selling its Alan Turing portrait, "AI God," for $1.3M — nearly 10x its estimated value. The artwork, created using Ai-Da’s robotic arms and camera-driven techniques, combines traditional and AI-based artistry and explores humanity’s relationship with technology.

• Anthropic has partnered with Palantir and AWS to bring its Claude AI models to U.S. intelligence and defense agencies, enabling advanced applications like data analysis, pattern recognition, and threat detection within classified environments. Integrated into Palantir’s high-security IL6 platform, the system will adhere to strict compliance protocols, avoiding controversial uses like weapons development.

• ByteDance unveiled X-Portrait 2, an AI system that animates static images by mapping realistic facial expressions and movements onto a driving video. Capable of handling subtle emotions and complex motions, the system works across realistic portraits and cartoon characters, with potential applications in animation, virtual agents, and TikTok integration.

• OpenAI CEO Sam Altman predicts artificial general intelligence (AGI) will be achieved by 2025, despite reports of slowing progress in large language models (LLMs). While Altman insists AGI requires engineering rather than scientific breakthroughs, OpenAI has formed a new "Foundations Team" to address challenges like limited high-quality training data.

• The Beatles’ AI-enhanced track "Now and Then" has made Grammy history, earning nominations for Record of the Year and Best Rock Performance. Using AI stem separation technology, the track isolated John Lennon’s vocals from a 1978 demo, enabling a polished final version.

• MIT researchers unveiled LucidSim, an AI system that trains four-legged robots in simulated environments, achieving 88% task success without real-world training. LucidSim uses ChatGPT to generate diverse training scenarios, combining physics and AI-generated imagery for tasks like obstacle navigation.

• Google DeepMind has open-sourced AlphaFold 3, its advanced protein prediction model, providing academic researchers with access to code and training weights for non-commercial use. Capable of mapping protein interactions with DNA, RNA, and drug compounds, the model has already mapped over 200M protein structures.

• Alibaba Cloud’s Qwen released its Qwen2.5-Coder series, a suite of open-source AI coding models, with the flagship 32B model rivaling GPT-4o and Claude 3.5 Sonnet in coding and reasoning tasks. Spanning six sizes and supporting over 40 programming languages, the models offer flexibility for diverse computing needs, with variants for fine-tuning or direct use.

• Japanese researchers have developed an AI system that detects conditions like high blood pressure and diabetes through short videos of a person's face and hands, with accuracy comparable to traditional methods. By analyzing subtle blood flow changes, the system achieves up to 94% accuracy and could be integrated into smartphones or smart mirrors.

• Apple is developing a wall-mounted AI smart home display designed as a central hub for tasks like video calls, appliance control, and media playback, expected to launch by March. The device will integrate Siri and Apple Intelligence for AI-driven interactivity, with a potential premium version featuring a robotic arm marketed as a home companion.

• Nous Research unveiled the Forge Reasoning API Beta, which boosts language model performance using advanced reasoning techniques like Monte Carlo Tree Search and Chain of Code. The system enables smaller models, like Nous’ 70B Hermes, to outperform giants such as o1 and Claude 3.5 Sonnet on complex tasks.

• Stanford researchers introduced Virtual Lab, an AI platform where specialized agents collaborate with human scientists, successfully designing nanobodies to combat COVID variants. The system uses AI "team members" with distinct specialties, coordinated by an AI Principal Investigator, to generate, refine, and validate research.

• OpenAI unveiled a sweeping AI infrastructure blueprint for the U.S., advocating for AI Economic Zones, a North American AI Alliance, and a modernized power grid to meet growing compute demands. The plan also includes a proposed $100B, 5-gigawatt data center to rival China’s advancements in AI infrastructure.

Quickfire News

• Microsoft began integrating Copilot AI features into standard Microsoft 365 subscriptions in select Asia-Pacific markets, hinting at a potential move away from the separate Copilot Pro subscription model.

• Black Forest Labs released an upgrade to its FLUX1.1 Pro model, introducing an "Ultra" mode for 4x higher image resolution in text-to-image generations and a "Raw" mode for more realistic outputs.

• Wendy’s partnered with Palantir to implement an AI-driven supply chain management system designed to predict shortages and automate inventory orders.

• Mistral launched a new multi-language content moderation API for its Le Chat platform, offering safety guardrails across nine policy categories for developers.

• Krea AI introduced custom model training capabilities, allowing users to develop personalized AI models tailored to specific characters, artistic styles, and product designs.

• Chinese EV maker XPENG unveiled Iron, a nearly 6-foot-tall robot with dexterous hands powered by its Turing AI chip, already deployed in its vehicle factories alongside autonomous driving technology.

• Nous Research debuted Nous Chat, a public chatbot interface powered by its Hermes 3-70B model.

• AI music generation startup Suno showcased demos of its upcoming v4 model, highlighting improved naturalness and consistency in audio samples.

• The U.S. Commerce Department ordered TSMC to halt exports of advanced AI chips to Chinese customers starting this week.

• Baidu is set to unveil AI-powered smart glasses with voice and camera capabilities at its Baidu World event, aiming to compete with Meta’s Ray-Ban smart glasses at a lower price point.

• A federal judge dismissed a copyright lawsuit filed by Raw Story and AlterNet against OpenAI, citing doubts about the news outlets' ability to prove harm from AI training data usage.

• The Washington Post launched "Ask The Post AI," a generative AI search tool leveraging the publication's archives to provide direct answers and curated results to reader questions.

• OpenAI VP of Research and Safety Lillian Weng announced her departure after seven years, marking another key leadership exit from the company.

• xAI introduced a free tier of its Grok chatbot in select regions, offering limited access to Grok 2, Grok 2 mini, and image analysis features.

• Lex Fridman released an interview with Anthropic CEO Dario Amodei, who discussed AI safety, predicted AGI could emerge by 2026-2027, and included conversations with Amanda Askell and Chris Olah.

• AI sales automation startup 11x raised $50 million in funding, bringing its valuation to $320 million as it expands its bots capable of handling sales tasks in 30 languages.

• Anthropic hired Kyle Fish as its first "AI welfare" researcher to examine whether future AI models might develop consciousness and require moral consideration.

• The Vatican partnered with Microsoft to create a digital AI-powered twin of St. Peter’s Basilica from 400,000 images, enabling virtual tours and aiding structural assessments ahead of the 2025 Jubilee.

• Jerry Garcia's estate teamed up with ElevenLabs to recreate the late musician’s voice using AI for audiobooks and written content in 32 languages.

• Leading AI companies are working on new benchmarks and testing methods, as current standards lag behind the sophistication of today’s AI models.

• Baidu announced new AI products at its Baidu World event, including the I-RAG text-to-image generator, the Miaoda no-code development tool, and AI-powered smart glasses.

• Alibaba introduced Accio, an AI-powered B2B search engine that uses natural language processing to connect global buyers and sellers, showing a 40% boost in purchasing intentions during pilot testing.

• Enterprise AI platform Writer raised $200 million in a Series C round, valuing the company at $1.9 billion, with plans to expand into healthcare, retail, and financial services workflows.

• OpenAI President Greg Brockman returned from a three-month sabbatical, taking on a new technical role within the company.

• Amazon launched a $110 million "Build on Trainium" initiative to support university AI research with free access to 40,000-chip clusters and open-source requirements for resulting innovations.

• Particle, an AI-powered news app, debuted on iOS, offering personalized summaries, multi-perspective analysis, and interactive features for better engagement with current events.

• Formation Bio, OpenAI, and Sanofi unveiled Muse, an AI system designed to accelerate clinical trial recruitment, which Sanofi is already using in Phase 3 trials to shorten drug development timelines.

• Chinese robotics firm Deep Robotics began commercial sales of its X30 quadruped robot, priced at $54,000, targeting industrial applications such as site inspections and security patrols.

• GEMA, a German performing rights organization, became the first to sue OpenAI over alleged copyright infringement of song lyrics, filing the case in Munich.

• AI safety advocate Dan Hendrycks joined Scale AI as an advisor, adding to his roles at The Center for AI Safety and xAI, with the $14 billion data labeling company focusing on safer AI systems.

• Microsoft launched adapted AI models, providing small, specialized language models tailored to address sector-specific challenges in manufacturing, automotive, and agriculture.

• DeepL introduced Voice, a real-time translation service supporting 13 spoken languages and 33 written languages, with an initial focus on text-based output for Teams meetings and in-person conversations.

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 ?