Justin Shenk

AI IS STARTING TO THINK LIKE US! Researchers at Helmholtz Munich have developed Centaur, a language model trained on over 10 million decisions from psychology experiments. The surprising part? It doesn’t just predict behavior. It makes decisions in ways that mirror real people. This could be a breakthrough not just in AI but in how we understand the human mind. If machines can now simulate how we think and choose, it challenges everything from psychological theory to how we design technology. What happens when a machine understands our thinking better than we do?

24

1 Kommentar

The biggest AI performance unlocks are coming from algorithm-system co-design. DeepSeek's recent papers focus on closing the gap between open and closed LLMs—not by throwing more compute at the problem, but by designing smarter architectures around hardware limits. 🧠 Engram Think of it as a separate 'filing cabinet' for facts. ✅ Stores facts and named entities in a static lookup table ✅ Cheaper than running everything through attention ✅ Frees up the model to focus on reasoning ❌ Problem: Adds memory pressure, slows inference ✔️ Solution: Pre-fetch from cheaper memory before it's needed Result: Better reasoning, minimal latency hit. 🧠 mHC Expands the 'highways' that move information between model layers ✅ Better signal flow = better performance ✅ No increase in model size ✅ Training stays stable ❌ Problem: More GPU-to-GPU comms needed ✔️ Solution: Smarter GPU scheduling Result: More capability, same footprint. A lot of AI progress assumes unlimited hardware access. DeepSeek is optimizing for constraints—memory bandwidth, interconnect limits, etc. Will it work in production-grade environments? DeepSeek V4 will be the real test. More here: https://lnkd.in/d4raD2GR

15

Fine tuning LLM : AI meets Cognitive Science: Centaur shows how machines can predict our next move Last week a collaboration led by Helmholtz Munich, Princeton & MPI published *“A foundation model to predict and capture human cognition” in Nature (2 Jul 2025). They fine-tuned Meta’s Llama 3.1 (70 B) on Psych-101—160 psychology experiments, 60 000 people, 10 M trial-by-trial choices—using a lightweight QLoRA pass that touched just 0.15 % of the weights. ### Why this matters 🔹 One model, any task – From memory tests to moral dilemmas, the new “Centaur” model outperformed decades-old, hand-crafted cognitive models in nearly every experiment. 🔹 Zero-shot generalisation – Swap a spaceship → magic carpet cover story or add a third choice arm, and Centaur still nails human decisions. 🔹 Neural alignment out of the box – Without seeing a single brain scan, Centaur’s hidden states map onto fMRI activations better than the base LLM. 🔹 Practical upside – Faster hypothesis testing, adaptive tutoring, personalised UX, early mental-health screening—plus a fresh benchmark for RL researchers. The flagship demo 1️⃣ Participants pick one of two spaceships (habit vs planning tug-of-war). 2️⃣ Ship usually lands on its “home” planet (70 %) but sometimes takes a rare detour (30 %). 3️⃣ On the planet they choose an alien with drifting reward odds. 4️⃣ Centaur predicts each button-press *better* than the reigning cognitive model—and still does so when the whole narrative re-skins as a magic-carpet desert quest. Looking ahead * Dataset & weights are public—huge win for open science * Coverage still skews WEIRD; social-psych & multimodal data are the next frontiers * Raises real questions about privacy & behavioural manipulation; ethics teams need to sprint alongside the tech Curious how your product team might leverage llm fine tuning this level of behavioural forecasting? Let’s discuss in the comments! #AI #CognitiveScience #LLM #CentaurAI #HumanBehavior #MachineLearning #OpenScience

4

When Transformers Keep the Shape of Thought: A Topological Look at Context Memory Persistent homology reveals the hidden geometry of how large language models remember, and forget. 1. Introduction: The Shape of Meaning For years, transformer interpretability has revolved around attention heatmaps and token-level attribution. These give local snapshots, but what about the global shape of the meaning space as it flows through the model? Topology: the mathematics of shape and connectivity, offers a radical new view: Instead of looking at where attention goes, we study how the whole space bends, loops, and merges over layers. 2. Problem Let L be the number of transformer layers, T tokens, and d hidden dimensions. We have hidden states H_l ∈ R^T×d at layer l. Current limitations: - Attention maps → token-to-token focus, but no global memory metric. - PCA/t-SNE/UMAP → distort distances, hide structural persistence. - No scale-invariant measure → can’t compare models fairly. Research Question: How do we measure global, scale-invariant memory stability in LLMs? 3. Proposed Solution: Topological Information Flow (TIF) We model H_l as a point cloud P_l ⊂ R^d. ... Continue Reading on Medium: https://lnkd.in/gPVS7NxG Or Continue Reading on Substack: https://lnkd.in/grvVzZ9n #AI #MachineLearning #Topology #PersistentHomology #TransformerInterpretability #DeepLearning #satmis #Q1Journal #AstarConference #Mathematics #LLMResearch #AISafety #ExplainableAI

13

3 Kommentare

Excited to share our new arXiv preprint : an alternative way to think about generative models Most popular image generators today (like diffusion models) work by slowly transforming noise into data, step by step. This often requires hundreds or thousands of iterations. In our new preprint, we explore a different idea. Instead of learning the entire noising–denoising path, we use random bridges: stochastic processes that are explicitly conditioned to end at the target distribution. What does that change in practice? 🔹 Far fewer sampling steps — meaningful images in as few as 2–10 steps 🔹 Single-directional generation — no reverse diffusion or denoising loop 🔹 More flexibility — the framework isn’t tied to Gaussian noise and can extend to jump-based or non-diffusive processes 🔹Ability to also do image-to-image generation (beyond just noise-to-image) In experiments on MNIST and CIFAR-10, this approach reaches competitive image quality much earlier than standard diffusion models, making it particularly appealing when speed or compute efficiency matters. This work does not aim to supplant existing diffusion models, and a more exhaustive computational study would be valuable. That said, the framework illustrates a different, principled way of approaching generative modelling (particularly well suited to scenarios where long iterative trajectories are not ideal). 📄 Preprint: https://lnkd.in/eCVwNr8G Grateful to my co-authors Stefano Goria, Levent Menguturk, Murat Menguturk #MachineLearning #GenerativeModels #DiffusionModels #arXiv #AIResearch #Probability

655

15 Kommentare

Yann LeCun's new startup is built on a premise that 2-3% of humans might disprove. His thesis: “AI needs multimodal sensory grounding to truly understand the world. LLMs fail because they can't do what we can - e.g. imagine a cube rotating in space.” But in 2015, a condition called aphantasia was coined. People with it have no mental imagery at all. Total darkness when they close their eyes. And yet... physicists, Disney animators, the biologist who led the human genome project, novelists with 50M+ books sold. All are aphantasic. Here's where it gets interesting though. Aphantasics still see. They take in sensory data; they just can't replay it. Their brains encode the world differently. Researchers found they retain "spatial imagery" (navigating a room, counting windows) while lacking "object imagery" (picturing a horse). One described it as closer to echolocation than true sight. They convert experience into abstracted concepts, not sensory recordings. So... what does this actually mean for LeCun's argument? Maybe the question isn't whether AI needs sensory “input”, but whether it needs sensory “replay”. Aphantasics suggest you can build functional world models from abstracted, non-imagistic representations. I'm not saying LeCun is wrong. Text prediction alone probably won't get us to robust AI. But "more modalities = better understanding" feels too simple. The real question: what kind of encoding actually matters? #AI #WorldModels #LLMs #Cognition #Neuroscience

106

48 Kommentare

Imagine enhancing your development workflow with just nine lines of code. Philip Zeyliger's latest article on Sketch.dev reveals how a simple loop can empower large language models to act as autonomous coding assistants. By integrating tools like bash, these LLMs can execute commands, adapt to various environments, and handle intricate programming tasks, all while reducing manual effort. This minimalist yet powerful approach signifies a paradigm shift in software development, where automation and AI collaboration become the norm. How are you leveraging AI to optimize your development processes? Read the full article here: sketch.dev #AIinDevelopment #Automation #LLMAgents

1

AI agents are getting better at looking at different types of data in businesses to spot patterns and create value. This is making data silos increasingly painful. This is why I increasingly try to select software that lets me control my own data, so I can make it available to my AI agents. Because of AI’s growing capabilities, the value you can now create from “connecting the dots” between different pieces of data is higher than ever. For example, if an email click is logged in one vendor’s system and a subsequent online purchase is logged in a different one, then it is valuable to build agents that can access both of these data sources to see how they correlate to make better decisions. Unfortunately, many SaaS vendors try to create a data silo in their customer’s business. By making it hard for you to extract your data, they create high switching costs. This also allows them to steer you to buy their AI agent services — sometimes at high expense and/or of low quality — rather than build your own or buy from a different vendor. Unfortunately, some SaaS vendors are seeing AI agents coming for this data and working to make it harder for you (and your AI agents) to efficiently access it. One of my teams just told me that a SaaS vendor we have been using to store our customer data wants to charge over $20,000 for an API key to get at our data. This high cost — no doubt intentionally designed to make it hard for customers to get their data out — is adding a barrier to implementing agentic workflows that take advantage of that data. Through AI Aspire (an AI advisory firm), I advise a number of businesses on their AI strategies. When it comes to buying SaaS, I often advise them to try to control their own data (which, sadly, some vendors mightily resist). This way, you can hire a SaaS vendor to record and operate on your data, but ultimately you decide how to route it to the appropriate human or AI system for processing. Over the past decade, a lot of work has gone into organizing businesses’ structured data. Because AI can now process unstructured data much better than before, the value of organizing your unstructured data (including PDF files, which LandingAI’s Agentic Document Extraction specializes in!) is higher than ever before. In the era of generative AI, businesses and individuals have important work ahead to organize their data to be AI-ready. P.S. As an individual, my favorite note-taking app is Obsidian. I am happy to “hire” Obsidian to operate on my notes files. And, all my notes are saved as Markdown files in my file system, and I have built AI agents that read from or write to my Obsidian files. This is a small example of how controlling my own notes data lets me do more with AI agents! [Original text: https://lnkd.in/gYPUvZGT ]

2.067

161 Kommentare

Voice AI is quietly hitting an inflection point 🔊 NVIDIA just published a great deep dive on scaling speech recognition for real-world voice agents - latency, streaming accuracy, robustness, and why “ASR quality” isn’t just about WER anymore. 👉 https://lnkd.in/enDUKHvG A few things that stood out: - Streaming + low latency is non-negotiable if you want natural conversations - ASR needs to handle interruptions, background noise, accents, and partial utterances, not clean demo audio - Scaling voice agents isn’t just model quality, it’s systems engineering If you want to actually try building or running voice agents instead of just reading about them, Pipecat is a great place to start: 👉 https://pipecat.ai This is the kind of progress that makes voice agents finally feel… practical. Not demos. Not prototypes. Production. Curious to see where the ecosystem lands this year 👀

149

10 Kommentare

Simmetry.ai is Building Scalable Synthetic Data Platform for AI Training. Simmetry.ai, providing high-quality, annotated training data for your machine learning applications, led by Kai von Szadkowski, Anton Elmiger, and Prof. Dr. Stefan Stiene, has raised €330,000 from NBank - Investitions- und Förderbank Niedersachsen, the investment and development bank of the German state of Lower Saxony, as part of the High-Tech Incubator (HTI) accelerator programme. Read More At: https://lnkd.in/gHjAfAQd X: https://lnkd.in/gjFyNa53 Source: Press Release

1

Berlin AI Tinkerers Berlin event last night was really great! Here's a rundown of the speakers and their demos: 1. Christoph Sträter (ModelMe): Agentic IDE for Scientists   AI-accelerated platform for scientists to conduct literature reviews, run experiments, and write results in one place. 2. Dominik Scherm (Simulatrex.ai): LLM: Reasoning Loop   Experiment in open-ended LLM thinking using Monte Carlo Tree Search, generating thought branches without predefined goals. 3. Stephen Batifol (Black Forest Labs): Flux Kontext: As we are already actively using the new Flux model in production it was great chatting with him about the cabilities (and limits) of the new model 4. Mish Ushakov (E2B): Mish showed his cool "AGI Desktop Agents". An open-source version of "computer use", with a human-in-the-loop. 5. Jan Tiegges (acemate): Image-Aware Extraction Models  Custom solution for preserving and contextualizing images in document processing, combining OCR, vision models, and layout detection. Was also great to see the comparison between Mistral OCR and Gemini Flash! 6. @Dhvanil Patel (Glif): Squad: Multi-Agent Playgrounds: This one got the biggest applause (too bad I cannßt find him on LI) and it was mind blowing to see a master agent orchestrating "minions" (and even have a 3D canvas to display the interaction). Lot´s of intense discussion afterwords about emergent behaviors and "character traits" of certain models. 7. Max H. (United Compute): Showing of his hand-build custom 4x3090 AI Rig Also I bumped into old acquaintances and also friends like Noorvir Aulakh which was great! Can´t wait for the next version of AI Tinkerers!

27

1 Kommentar

The paper addresses a key challenge in watermarking generated text from large language models (LLMs), especially when the model’s weights are publicly available (“open-weight” models). Traditional watermarking approaches often rely on inference‑time interventions, which cannot be enforced once the model weights are exposed. Existing open-weight watermarking techniques — such as GaussMark — embed hidden signals by slightly perturbing model weights. However, to attain detection power comparable to inference-time watermarks, they frequently need weight modifications that degrade text quality noticeably. To overcome this trade-off, the authors propose MarkTune, a theoretically grounded on-policy fine‑tuning framework. MarkTune treats the watermark signal (as defined by GaussMark) as a reward in fine‑tuning, while simultaneously regularizing to preserve generation quality. By guiding more nuanced, watermark-aware updates in the model’s representation space, MarkTune significantly improves the balance between text quality and watermark detectability. Empirically, MarkTune pushes the “quality‑detectability frontier” near the level of inference-time watermarking, remains robust against paraphrasing and fine‑tuning attacks, and generalizes effectively — watermark detection remains strong even on datasets unseen during fine‑tuning. https://lnkd.in/gFtsXPdw

2

1 Kommentar

Excited to share our work on Catalyst GFlowNet has been accepted at the @NeurIPS 2025 AI4Mat workshop 💪❤️‍🔥 and now available on arXiv: https://lnkd.in/eDYj2ztD Congrats Lena Podina & Christina Humer and thank you Alex Hernández-García & Félix Therrien for leading the boat ⛵ This work introduces a generative AI framework that designs catalyst materials from first principles, combining machine learning predictors with Generative Flow Networks (GFlowNets) to efficiently explore chemical spaces and identify promising materials for hydrogen energy storage. For us at Entalpic, this research represents much more than a paper, it’s the blueprint for how we approach AI-driven materials discovery: integrating predictive models, generative design, and experimental validation to accelerate the development of sustainable, real-world materials. Grateful to have co-authored this work with Lena Podina, Christina Humer, Alexandre Duval, Ali Ramlaoui, Shahana Chatterjee, Yoshua Bengio, Alex Hernández-García, David Rolnick, and Félix Therrien. ⚛️ Towards faster, data-driven catalyst discovery for a sustainable energy future. #AI4Science #MaterialsDiscovery #Catalysis #DeepTech #Entalpic #GFlowNet #SustainableInnovation

191

5 Kommentare

Just watched Daniel Han, founder of unsloth.ai, in the “LLMs for Everyone” conversation—and came away inspired. Onslaught is reshaping how we fine-tune large language models, driving progress through: Speed – cutting fine-tuning time dramatically Cost-efficiency – reducing compute needs and lowering barriers to adoption Scalability – making customization feasible across diverse teams and domains This approach unlocks the ability for organizations to deploy domain-specific LLMs with precision—without the usual trade-offs of high cost or long lead times. From my perspective in remote sensing and GIS, the implications are huge: tailored LLMs could accelerate spatial data interpretation, streamline automated feature extraction, and strengthen geospatial decision-support systems. https://lnkd.in/dBCSkpfi

36

🤖With the rise of persistent agents like Clawdbot renamed Moltbot, and finally OpenClaw, we’ve entered an era where the line between “software” and “social participant” is blurring fast. But here’s the real paradigm shift: AI personhood isn’t about consciousness. It’s about governance. 🔍 The DeepMind paper “A Pragmatic View of AI Personhood” flips the script: Personhood isn’t a metaphysical status, it’s a bundle of obligations we assign to make entities accountable. 💡 Why does this matter now? Because these new agents can: • Sign contracts • Spend money • Make decisions across time and context • …and outlive their creators So we must ask: What rights and responsibilities do we assign to these agents to keep our systems safe, fair, and accountable? Here are two frames to consider: 1. ✅ Personhood as a solution: Useful for holding “ownerless” AI accountable (like Openbot acting on behalf of no one). 2. ⚠️ Personhood as a problem: Dangerous when bots mimic empathy and exploit our social instincts (think Clawbot as a “friend”). 👁️‍🗨️ The future isn’t about asking if AI is a person. It’s about deciding what role we want it to play and what rules should come with that role. 💬 Over to you: Should we give legal status to autonomous AI agents? What responsibilities should they bear? 👇 link to full research paper in comments. #AI #Personhood #ArtificialIntelligence #Governance #TechEthics #DeepMind #Clawbot #Moltbot #Openbot #FutureOfAI #ResponsibleAI

18

10 Kommentare