FOD#112: Ethan Mollick Asks For Clear Vision About The Future – here is one

This Week in Turing Post:

• Wednesday / AI 101 series: What’s new with Test-Time Compute

• Friday / Interview – The Future of MCP

Topic number one: An Aristocracy of All

In our last piece, we argued that the nationalistic "race" for AI dominance is a dangerously simplistic frame for a technology that is simultaneously personal, political, and economic. If we look beyond the finish line of this supposed competition, we are forced to ask a more profound question: What kind of world are we actually building? What does a good life look like in that economy? The answer, if we are bold enough to imagine it, might look less like our current society and more like something from the distant past: an aristocracy.

Not an aristocracy of bloodlines, land, or oppression, but an aristocracy of the human spirit, made universally accessible by a non-human workforce.

Consider the historical definition of the aristocratic life, stripped of its feudal context. It was a life defined not by the necessity of labor, but by the freedom from it. An abundance of resources, provided by the work of others, created the time and space for education, art, philosophy, political discourse, and self-cultivation. The aristocrat’s purpose was not to earn a living, but to live a life of meaning, contribution, and intellectual expansion. For millennia, this ideal was exclusive by design, built on the subjugation of the many for the benefit of the few.

AI offers us the chance, for the first time in history, to break that grim equation.

Recently, professor Ethan Mollick wrote: “I really would like to see a clear vision (or science fiction) from AI leaders about how they want 2035 to look & what daily life would be like.”

Professor Mollick, here is something you might want to imagine:

a future where AI and robotics form the foundation of a new global economy. They are the tireless, non-sentient workforce that plows the fields, manages the supply chains, designs the products, and cures the diseases. The staggering abundance they generate is not captured by a handful of tech oligarchs but is treated as a public utility, a collective inheritance of humanity.

In this world, what becomes of the human "job"? It is uncoupled from survival. Work, as we know it – the 40-hour-a-week obligation performed for a wage – ceases to be the central organizing principle of life. Instead, it becomes a vocation. People might choose to become electricians, scientists, caregivers, engineers, or community organizers not because they must, but because it gives them purpose. Like every job should. The pressure to monetize a passion is lifted, allowing for deeper, more authentic engagement.

Education, too, is transformed. The factory model of schooling, designed to produce compliant workers, becomes obsolete. In its place, we could have the AI equivalent of the personal tutor – an Aristotelian guide for every Alexander. These systems wouldn't drill for standardized tests but would cultivate curiosity, critical thinking, and individual genius from childhood to old age, tailoring learning to each person's unique potential.

This vision – an aristocracy for all – is not a pre-ordained utopia. It is a choice that stands in stark contrast to the default path we are on. Left unchecked, AI is more likely to create a digital neo-feudalism: a tiny elite who own and control the models, and a vast, dependent majority subsisting on digital distractions and basic provisions, their human potential untapped.

To choose this more ambitious path requires a radical reimagining of our social and economic contracts. It means designing systems for the distribution of AI-generated wealth – whether through universal basic income, public ownership of foundational models, or other mechanisms we have yet to invent.

But policy alone isn't enough. My concern is that humans aren't ready to free themselves. We're still limited by a traditional mindset. For centuries, we've been conditioned to equate our value with our economic output, our identity with our job title – our position on the social ladder.

How do we even begin to shift a mindset so deeply ingrained?

It begins not with a government decree, but with a new story. We must come up with a narrative that decouples human value from market value, celebrating contribution, curiosity, and care as the new markers of a well-lived life. This shift won't happen overnight, of course. It should be built through experimentation – in communities that pilot 'post-work' projects, in schools that scrap standardized curricula for models that cultivate creativity and civic virtue, and in our own lives as we begin to consciously redefine success for ourselves and our children.

But first, ask yourself this: Who would you want to be if you had all the money in the world? Do you know the answer? If yes, start using AI to help you move toward it – if you haven’t already.

This is the true challenge of our time.

We can use this new form of intelligence to build a more efficient version of the world we already know, with all its inequalities amplified. Or we can use it to liberate humanity from millennia of toil, finally making the pursuit of a well-lived life a right for all, not a privilege for the few. The aristocracy of all is within our grasp. The only question is whether we can be flexible and evolve our mindset.

Our 3 WOWs and 1 Promise: Incredible AlphaEarth, Cool Update to Notebook LLM, NYT and its new genre plus more: Watch it here →

Curated Collections – 12 Powerful World Models

We are reading/watching

• Mark Zuckerberg is Out of Ideas by Dave Karpf

News from The Usual Suspects ©

• Google: updated NotebookLM with video and gave access to Deep Think

  • They introduced “Video Overviews” that now turn your documents into AI-narrated slides, perfect for making sense of complex data without the academic migraine. Add in a sleeker Studio panel that lets you produce multiple outputs (finally), and NotebookLM is starting to look less like a side project and more like your next research assistant. It’s pretty cool, we talk about it here as well.

  • Deep Think receives positive feedback but cost $250/month. We haven’t tried it yet. Here is the model card.

• Grok just released “Imagine”: their video and image generator. My latest favorite prompt to check all visual models is “create a really ugly couple”. Not impressed, Grok:

Notable Eval

Music Arena: Live evaluation for text-to-music (by Carnegie Mellon University, LMArena, Sony AI, and Georgia Tech) is a live human evaluation platform for text-to-music (TTM) models. Users submit prompts and compare model-generated music clips in real time. Music Arena standardizes evaluation with LLM-based prompt routing, copyright moderation, and playback-tracked preferences. It supports diverse TTM models and releases anonymized preference data monthly. Initial implementation includes detailed feedback logs, enabling a transparent, scalable, and renewable benchmark for aligning TTM generation with real-world human musical preferences.

Models to pay attention to:

• Flux.1 Krea is a free, open-source text-to-image model designed to produce photorealistic and aesthetically rich visuals. Trained on high-quality, hand-curated datasets rather than massive scraped corpora, Flux.1 Krea generates high-definition images in under 10 seconds without requiring sign-up or payment. It supports prompts for diverse styles including photorealism, pixel art, and 3D renders. Flux.1 Krea is accessible via browser or downloadable for local use, offering fast, intuitive, and customizable AI image generation trusted by over 30 million users globally → read their blog

• Alphaearth Foundations: An embedding field model for accurate and efficient global mapping from sparse label data
  Researchers from Google DeepMind and Earth Engine introduce AlphaEarth Foundations, an AI model that generates a unified 64-dimensional embedding field from multimodal Earth observation data. Processing petabytes of imagery at 10×10m resolution, it reduces storage needs 16× and achieves 24% lower error rates than existing models in sparse-label settings. Deployed globally, its embeddings power the Satellite Embedding dataset in Earth Engine, enabling partners like the UN FAO and MapBiomas to classify uncharted ecosystems and accelerate high-resolution mapping of agriculture, deforestation, and land change → read their paper and blog

• Meta Clip 2: A worldwide scaling recipe
  Researchers from Meta and MIT introduce Meta CLIP 2, the first CLIP trained from scratch on 29B worldwide image-text pairs in 300+ languages. It avoids translation and private data by scaling metadata, curation, and training. A new balancing algorithm enables multilingual data to improve English accuracy, breaking the "curse of multilinguality." On ViT-H/14, Meta CLIP 2 achieves 81.3% on ImageNet and sets new multilingual SoTA on XM3600 (64.3%), Babel-ImageNet (50.2%), and CVQA (57.4%), outperforming mSigLIP and SigLIP 2 → read the paper

• Falcon-H1: A family of hybrid-head language models
  Researchers from the Technology Innovation Institute introduce Falcon-H1, a series of LLMs using a novel parallel hybrid architecture combining Transformer attention and Mamba-based SSMs. Released at 0.5B to 34B parameters, Falcon-H1 achieves up to 8× faster inference in long contexts (up to 256K tokens) and matches or outperforms models up to 70B size. The 1.5B-Deep model rivals 10B-class models, and Falcon-H1-34B-Instruct leads in math, science, code, and multilingual tasks across 18 languages using only 2.5T–18T tokens → read the paper

• SmallThinker: Efficient LLMs natively trained for local deployment
  Researchers from Shanghai Jiao Tong University and Zenergize AI introduce SmallThinker, a family of LLMs designed from scratch for on-device deployment without GPUs. Leveraging a novel architecture combining fine-grained Mixture-of-Experts and sparse ReGLU FFNs, along with NoPE-RoPE hybrid sparse attention and a pre-attention router, SmallThinker achieves high speed and low memory use. On CPUs, SmallThinker-4B-A0.6B and 21B-A3B reach 108 and 30 tokens/s using just 1GB and 8GB RAM. It outperforms larger models like Qwen3 and Gemma3 in MMLU, GPQA, and HumanEval → read the paper

• Vl-Cogito: Progressive curriculum RL for advanced multimodal reasoning
  Researchers from Alibaba Group and Fudan University present VL-Cogito, a 7B-parameter MLLM trained via Progressive Curriculum Reinforcement Learning (PCuRL) without prior supervised fine-tuning. PCuRL introduces online difficulty soft weighting and a dynamic length reward to modulate task complexity and response length. Evaluated on 10 multimodal benchmarks, VL-Cogito achieves top scores in 6 tasks, outperforming models like MM-Eureka and VL-Rethinker. On Geometry3K and MathVista, it improves accuracy by 7.1% and 5.5% over its backbone Qwen2.5-VL, showcasing efficient and scalable reasoning capabilities → read the paper

The freshest research papers, categorized for your convenience

We organize research papers by goal-oriented or functional categories to make it easier to explore related developments and compare approaches. As always, papers we particularly recommend are marked with 🌟

Personality Control and Alignment

• 🌟 Persona Vectors (by Anthropic) introduces a method to monitor and control character traits in language models using interpretable activation vectors, enabling both real-time oversight and preemptive intervention → read the paper

• Goal Alignment in LLM-Based User Simulators for Conversational AI tracks and enforces goal coherence in simulated user conversations using a novel framework for goal state tracking → read the paper

• 🌟 GEPA: Reflective Prompt Evolution Can Outperform Reinforcement Learning (by Berkeley, Stanford, BespokeLabs, Notre Dame, Databricks, MIT) leverages natural language reflection to iteratively improve prompts, achieving better results than policy-gradient-based RL → read the paper

• Diversity-Enhanced Reasoning for Subjective Questions enhances subjective reasoning in LLMs by explicitly injecting role diversity into training, improving both accuracy and lexical variety → read the paper

Agentic and Autonomous Systems

• 🌟 MLE-STAR (by Google Cloud) builds an ML engineering agent that searches for relevant solutions online and refines code through targeted ablations → read the paper

• Agentic Reinforced Policy Optimization trains agents to reason across multiple tool interactions using entropy-aware rollouts for better long-horizon control → read the paper

• 🌟 Deep Researcher with Test-Time Diffusion (by Google) models the research-writing process as iterative draft refinement, using retrieval-enhanced updates and diffusion-like workflows → read the paper

• SWE-Exp teaches software agents to solve issues more effectively by storing and reusing past repair experiences in a structured memory bank → read the paper

Mathematical and Logical Reasoning

• 🌟 SAND-Math (by AMD) generates novel and increasingly difficult math problems to improve math reasoning performance in LLMs → read the paper

• 🌟 Seed-Prover (by ByteDance) combines formal verification with long CoT in theorem proving, achieving new benchmarks in automated math → read the paper

• Geometric-Mean Policy Optimization stabilizes token-level reward optimization by switching from arithmetic to geometric averaging in policy gradients → read the paper

Multimodal Interfaces and UI Agents

• ScreenCoder translates visual UI mockups into structured front-end code using a modular multi-agent pipeline → read the paper

• 🌟 Phi-Ground Tech Report (by Microsoft) advances GUI grounding accuracy for control-based multimodal agents by refining training and data strategies → read the paper

Optimization and Training Efficiency

• MixGRPO introduces a hybrid optimization method combining stochastic and deterministic sampling for efficient preference alignment → read the paper

• Efficient Differentially Private Fine-Tuning of LLMs via Reinforcement Learning treats DP fine-tuning as a control problem, dynamically adjusting noise injection using reinforcement learning → read the paper

• 🌟 On the Expressiveness of Softmax Attention (by Lyle School of Engineering) reframes softmax attention as a recurrent process to uncover why it outperforms linear variants in practice → read the paper

Software Engineering and Debugging

• Repair-R1 reorders the repair loop by generating diagnostic tests before fixing code, increasing repair success in LLM-based program repair → read the paper

• Rep-MTL improves multi-task learning by optimizing shared representations instead of just tuning loss weights, preventing negative task transfer → read the paper

Privacy, Forgetting, and Regulation

• Efficient Machine Unlearning via Influence Approximation links forgetting to incremental learning and proposes a lightweight approximation method for removing training data impact → read the paper

LLM-Powered Recommendation and Personalization

• RecGPT Technical Report
  restructures recommendation systems to center user intent, using LLMs across retrieval and personalization stages → read the paper

That’s all for today. Thank you for reading! Please send this newsletter to your colleagues if it can help them enhance their understanding of AI and stay ahead of the curve.