The Illusion of Stagnation

Good morning,

Something strange is happening.

While researching this week's newsletter, my first reaction was one of underwhelm. Nothing spectacular this week. Not like last week’s Magnus release.

Sure, an AI model scored very high on a “frontier math” test.

Meh.

Lame.

I want MORE progress. Give me MORE !

But upon closer examination, I realized my underwhelm is the consequence of my own ignorance. I used to have an aversion to math so the magnitude of this breakthrough is just incomprehensible to me. And “frontier math” is something that goes above almost everyone’s head.

It’s like when a few years ago AlphaGO beat the best player in the world in Go, by playing a move that seemed outrageous. A move that was completely unexpected and seemed like a huge mistake at first. Everybody thought AlphaGO had blundered.

Go experts were already gloating. The game of Go would prove to be unbeatable by AI -just like they predicted.

It was only 30 moves later that it slowly dawned on all the onlookers and experts that this seemingly erratic move was nothing but pure genius.

The human player was blindsided and the AI won by deploying a tactic noone had ever seen.

And herein lies my point : AI is advancing faster than most of us realize. And it’s starting to do it in ways that we cannot wrap our heads around. Because it is doing it in fields we barely understand ourselves anymore.

Which is in itself difficult to wrap your head around (very meta - I know).

You know - when I became a freelancer and had a first meeting with an accountant he was talking to me about all kinds of things related to accounting. And he might as well have been speaking in Mandarin Chinese - i felt completely lost.

And that’s not stupidity (I hope) but it’s simply someone who is doing something professionally that is completely different to what I do and it uses all kinds of jargon I’m not familiar with.

But you can have the same type of experience within your own field of expertise. Where you meet someone that is so good at a certain aspect of it - you can also feel like you know nothing.

But with generative AI - it’s not only active in different fields - it is starting to operate in realms where usually extremely high IQ individuals dwell. Above our heads - so to speak.

And where progress becomes difficult to perceive. And it’s all happening faster than we realize.

Luckily this week, research showed us that AI has it’s own “Moore’s Law” and progress can be measured.

If I'd told you back in 2022 that Artificial General Intelligence (AGI)—machines with human-level thinking capabilities—would arrive before 2030, you'd probably have thought I was losing my grip on reality. It sounded like pure science fiction then. But now, respected AI experts agree that AGI could realistically emerge as soon as next year and certainly within the decade.

Initially, I was skeptical. After all, AI hype often fuels financial speculation as much as genuine progress. But when even AI researchers—sometimes anonymously—voice their concern on platforms like X or Reddit that we're moving dangerously fast, it's hard to dismiss entirely. In Silicon Valley, the consensus is clear: something unprecedented is just around the corner.

Yet strangely, to those of us using everyday tools like ChatGPT, it feels like progress has stalled. That's because we aren't using AI to unravel cosmic mysteries or tackle complex mathematics. We're just drafting emails, conducting quick research, or maybe generating ad copy. Our interactions create the illusion of stagnation because the real leaps are now happening beyond our own cognitive reach.

This is why last week's release of Manus resonated so strongly. Manus isn't fundamentally groundbreaking—it's essentially a better interface, unlocking existing potential within AI systems. But to us, it still felt like a significant leap forward.

Meanwhile, most people—including governments and institutions—are alarmingly unprepared. There's no clear roadmap for managing the economic upheaval or potential risks AI will inevitably create. Instead, we're standing on the brink of massive transformation, hoping to figure it out as we go.

AI is jumping forward dramatically. Models that struggled with simple arithmetic a few years ago now effortlessly solve advanced math. Entire industries are reshaping overnight as companies deeply integrate AI. Programmers increasingly act as supervisors, guiding AI-generated software rather than writing code themselves.

It's reminiscent of that Netflix movie "Don’t Look Up," where people ignored an incoming asteroid, too distracted by social media and trivial issues. Are we similarly distracted right now?

What do you think? Are we closer to AGI than most of us realize?

Until next week,

Stay curious,

Welcome to the Blacklynx Brief!

AI News

• OpenAI has submitted a 15-page proposal to the White House’s AI Action Plan, advocating for federal oversight while seeking to block state-level AI regulations. The company warns that the 781 state AI bills introduced this year could harm U.S. innovation against China and calls for expanded copyright protections, infrastructure investment, and access to government datasets. OpenAI also urged the U.S. to ban DeepSeek, calling it a “state-controlled” security risk.

• Cohere has launched Command A, an enterprise AI model that delivers top-tier performance while running on just two GPUs, making it significantly faster and more cost-efficient than GPT-4o and DeepSeek-V3. With a 256K context window, multilingual capabilities, and advanced retrieval-augmented generation (RAG) features, Command A is designed for private enterprise deployments.

• Google has introduced new personalization features for Gemini AI, allowing the assistant to access users' Search history—and eventually Google Photos and YouTube—to provide more tailored responses. The feature, powered by Gemini 2.0 Flash, is opt-in and restricted to users over 18, ensuring control over data usage.

• Baidu has launched ERNIE 4.5 and ERNIE X1, two aggressively priced multimodal AI models designed to compete with Western leaders at a fraction of the cost. ERNIE 4.5 improves language skills, reasoning, and coding while costing just 1% of GPT-4o's price at $0.55/$2.20 per million tokens, while ERNIE X1 matches DeepSeek’s R1 at half the cost. As China pushes AI prices lower, Western companies may be forced to respond, potentially triggering a global AI price war.

• A federal judge has denied Elon Musk’s request for a preliminary injunction against OpenAI, while fast-tracking the trial and dismissing some of Musk’s claims. Court filings allege Musk once wanted to merge OpenAI into Tesla as a for-profit entity, contradicting his lawsuit’s claims that OpenAI has abandoned its nonprofit mission. With OpenAI’s rumored $40B SoftBank investment hinging on its for-profit pivot, the lawsuit’s outcome could significantly impact its future and the AI industry at large.

• Harvard and MIT researchers have introduced TxAgent, an AI system that provides real-time, personalized treatment recommendations by analyzing drug interactions, patient history, and biomedical data. Using 211 specialized tools and sources like openFDA, it continuously refines recommendations through structured reasoning and function calls.

• Roblox has unveiled Cube 3D, an open-source AI tool that generates 3D objects from text prompts, making game asset creation faster and more accessible. Unlike traditional methods, Cube 3D trains on native 3D data and uses 3D tokenization for shape prediction, with future plans for 4D scene generation. Combined with AI-driven updates to Roblox Studio, these tools lower the barrier to game development and supercharge user creativity on the platform.

• Zoom has upgraded its AI Companion with agentic capabilities, allowing it to detect and complete tasks like scheduling, transcription, and document generation across the platform. New features include Zoom Tasks, AI-powered calendar management, and a $12/month Custom AI Companion add-on with personalized AI coaching and avatars. As Zoom shifts to an AI-first strategy, these updates move it closer to CEO Eric Yuan’s vision of AI digital twins handling meetings autonomously.

• Google Research and Muon Space have launched FireSat, an AI-powered satellite capable of detecting wildfires as small as 5x5 meters within minutes. Using infrared sensors and AI analysis, FireSat significantly improves over current low-resolution satellite imagery, with plans to expand to a 50-satellite constellation for near-global coverage. As wildfires intensify worldwide, this system could revolutionize early detection, helping emergency responders contain fires before they become disasters.

• Nvidia CEO Jensen Huang kicked off GTC 2025 with a keynote packed with AI and chip advancements, calling the event “AI’s Super Bowl.” Key announcements included the Blackwell Ultra GPU (late 2025), Vera Rubin (2026), and Feynman (2028), alongside the Isaac GR00T N1 humanoid robot foundation model and a Newton robotics physics engine built with DeepMind and Disney. Nvidia also unveiled DGX Spark and DGX Station for AI workstations and announced a self-driving car partnership with GM, reinforcing that AI’s acceleration is far from slowing down.

• Anthropic is reportedly developing a voice mode for Claude while doubling down on enterprise AI, shifting focus from mass adoption to high-value business users. CPO Mike Krieger hinted at upcoming features for meeting analysis, report generation, and workflow automation, with Claude potentially integrating with Amazon and ElevenLabs for voice AI. By prioritizing corporate productivity tools over consumer AI, Anthropic is carving out a niche where companies are willing to pay top dollar for efficiency gains.

• Researchers at METR have identified a "Moore's Law" for AI capabilities, showing that the length of tasks AI agents can complete autonomously has doubled every 7 months since 2019. Current top models like Claude 3.7 Sonnet can handle tasks requiring 59 minutes of human effort with 50% reliability, while GPT-4 is limited to 8–15 minutes. If this trend continues, AI could autonomously complete month-long projects by 2030, marking a major shift in automation and workforce dynamics.

• Over 400 Hollywood creatives, including Ben Stiller, Paul McCartney, and Cate Blanchett, have signed an open letter urging the Trump administration to reject OpenAI and Google’s proposals to expand AI training on copyrighted works. OpenAI framed exemptions as a national security issue, while Google argued fair use already supports AI development, but critics insist AI firms should license content like any other industry. This clash highlights the growing legal and ethical tensions between AI innovation and traditional intellectual property rights.

• Nvidia has unveiled its Llama Nemotron family of open-source reasoning models, optimized for agentic AI and complex decision-making. Available in Nano (8B), Super (49B), and Ultra (249B) sizes, early benchmarks show Super outperforming Llama 3.3 and DeepSeek V1 across STEM and tool-based tasks. With an AI-Q Blueprint framework launching in April, Nvidia is positioning itself as a full-stack AI powerhouse—spanning everything from cutting-edge hardware to high-quality reasoning models.

Quickfire News

• AI2 released OLMo 2 32B, the first fully open model to outperform GPT-3.5 and GPT-4o mini on academic benchmarks while using only one-third of the training compute of comparable models like Qwen 2.5 32B.

• Microsoft and Xbox introduced "Copilot for Gaming," an AI assistant designed to help players get into games faster, provide in-game coaching, and enhance social experiences, with early access launching on mobile soon.

• Alibaba launched New Quark, a redesigned AI assistant app now powered by its Qwen reasoning model.

• Google added YouTube support to the Gemini API and Google AI Studio, enabling the model to interact with video content using vision capabilities.

• Former OpenAI researcher Andrej Karpathy predicted that AI will reshape web content, with pages optimized for LLMs and context windows rather than human readability.

• Insilico Medicine secured $110M to expand its AI-powered drug discovery platform, featuring multimodal foundation models and a fully automated robotic lab with a bipedal humanoid AI Scientist.

• Google’s Gemini 2.0 Flash is reportedly being misused to remove watermarks from images, including those from Getty and other platforms.

• Figure launched BotQ, a humanoid robot manufacturing facility capable of producing 12,000 units per year, with plans to scale to 100,000.

• Patronus AI introduced the industry’s first multimodal LLM-as-a-judge, designed to help developers detect and mitigate reliability issues in multimodal AI apps.

• Pika Labs released 16 new effects for its AI video platform, enabling users to morph images into different character videos.

• Sesame, known for its realistic AI voice tech, open-sourced its Conversation Speech Model (CSM-1B) for text-to-speech tasks.

• Vogent AI launched self-improving voice agents that learn from real failures and require no prompt engineering.

• Y Combinator CEO reported that nearly 25% of their startups now have 95% of their code written by AI, allowing small teams to hit $10M revenue with fewer engineers.

• OpenAI CPO Kevin Weil predicted 2025 will be the year AI permanently surpasses humans in programming, calling it a "democratizing effect" that will allow anyone to create software.

• Mistral AI released Small 3.1, an open-source multimodal model with a 128K token context window, outperforming Gemma 3 and GPT-4o Mini on key benchmarks.

• xAI acquired generative video startup Hotshot, shutting down new video creation to integrate and scale training on xAI’s Colossus cluster.

• Chinese researchers introduced ReCamMaster, an AI system that can edit camera angles and movement in a video while preserving original scene details.

• MagicLab showcased its Magicbot humanoid running continuously outdoors for four minutes, as it prepares for an upcoming half marathon in Beijing.

• Google partnered with MediaTek to develop next-gen Tensor Processing Units, reducing its reliance on Broadcom and expanding in-house chip development.

• Perplexity released a new commercial featuring Emmy-winning actor Lee Jung-jae, directly challenging Google Search in a Squid Game-style ad.

• Google introduced Canvas, a collaborative space in Gemini for document editing and code creation, along with the addition of Audio Overviews to the Gemini platform.

• Meta announced that Llama has officially surpassed 1 billion downloads, up from 650M in December 2024, marking a major milestone for its open-source AI model.

• OpenAI VP of Research Liam Fedus is departing to launch an AI materials science startup, with OpenAI planning to invest in and partner with the venture.

• Google revealed TxGemma, a Gemma-based AI model collection aimed at accelerating drug discovery, set to launch later this month.

• Tencent’s Hunyuan released 3D 2.0 MV and 3D 2.0 Mini, two new 3D generation models for high-quality multiview shape creation.

• Stability AI unveiled Stable Virtual Camera, a diffusion model that converts single images into 3D videos with 14 dynamic camera path options.

• Anthropic-backed Graphite launched Diamond, an AI code review tool that provides codebase-aware feedback and fixes, alongside a $52M Series B funding round.

• Google AI and UC Berkeley researchers proposed "inference-time search" as a new AI scaling method, generating multiple answers in parallel and selecting the best one.

• LG launched EXAONE Deep, a 32B parameter reasoning AI that matches DeepSeek V1 in math, science, and coding performance.

• Muse introduced Muse S Athena, a wearable headband that uses EEG and oxygen level sensors for AI-powered cognitive fitness training.

• Nvidia and xAI joined the AI Infrastructure Partnership, alongside Microsoft, BlackRock, and MGX, aiming to raise $30B initially and up to $100B for AI data centers.

• xAI debuted its first image generation API, powered by ‘grok-2-image-1212’, allowing developers to generate multiple JPG images per request at $0.07 each.

• Microsoft partnered with neuroscience AI startup Inait to develop brain-inspired AI that learns from real-world experiences instead of traditional data patterns.

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 !