Joe Hellerstein

A team of UC San Diego scholars just published a bold claim in Nature. They argue today's large language models have already reached artificial general intelligence. Here's their evidence. Philosopher Eddy Keming Chen, AI professor Mikhail Belkin, linguist‑computer scientist Leon Bergen, and data scientist David Danks make their case. They propose a tiered framework for judging intelligence. Their core argument is simple. General intelligence doesn't require perfection. It requires flexibility across domains. They point to a 2025 study where GPT‑4.5 was judged human‑like 73% of the time in a Turing‑test‑style evaluation. Humans also make mistakes and 'hallucinate' facts. So why should we hold machines to a higher standard? The authors differentiate between: 🔹 General intelligence (flexible problem-solving) 🔹 Superintelligence (far surpassing human capability) They claim current frontier models satisfy the first two tiers of their evidence framework. This includes basic conversation and complex reasoning. Beyond the technical debate, they address the emotional resistance. Acknowledging machine intelligence can feel like a paradigm shift. It challenges our place in the world. They warn that industry pressures for speed and profit can obscure genuine assessment. The call is for interdisciplinary collaboration and ethical governance. This isn't just about benchmarks. It's about how we define intelligence itself. The paper invites the AI community to move past denial. It asks for 'compassionate curiosity' about non‑human minds. What's your take? Does flexible problem-solving across domains equal general intelligence, or are we missing a key ingredient? #AGI #ArtificialIntelligence #LLM #AIResearch #FutureOfWork 𝗦𝗼𝘂𝗿𝗰𝗲꞉ https://lnkd.in/gfgwqSCj

5

1 Comment

Session date: April 10, 2026, noon CST Math Challenge: Can a graph have more vertices pointed at than pointing? Agent Skill with Anthropic AI News for this week, especially Anthropic Mythos

2

1 Comment

Last week, dbt Labs launched the public beta of dbt Fusion, a Rust based engine rewrite that brings a series of neat new features to our favorite data transformation tool: 🛠️ Real-time SQL validation and autocompletion 👀 Live previews of CTEs and compiled code 🔍 Rich model lineage at the colomn and table levels ⚡️ 30x faster project parse time (thanks Rust 🦀) 💻 Deep VS Code / Cursor integration At Opsima, dbt is a key component of our data stack (along with Prefect and Snowflake). dbt powers all data transformations across our cost and usage models and commitment automation engine. And being Cursor power users, the idea of a faster, smarter model development workflow within our IDE is quite exciting! That said, the reaction hasn’t been all positive. Reddit’s on fire with concerns about what Fusion means for the future of dbt Core and open source. https://lnkd.in/e9WSFARf dbt Labs insists Core stays under the Apache 2.0 license and fully supported, but Fusion is only “source-available”, not entirely open source, and licensed to limit hosting it as a service which raises some concerns. https://lnkd.in/ekVzgxcW For teams like ours though, we’re confident Fusion will help us ship faster and smarter while keeping core flows portable. What is your own take on the launch? #dbt #dbtFusion #FinOps #DataEngineering #OpenSource

13

The paper introduces H-AIM, a hierarchical planning framework designed to enable teams of heterogeneous robots to carry out complex, long-horizon tasks from high-level natural language instructions. To bridge the gap between intuitive task descriptions and executable robot actions, H-AIM uses a large language model (LLM) to interpret instructions and generate formal Planning Domain Definition Language (PDDL) problem descriptions. These descriptions are then fed into a classical planner to derive optimized action sequences. Finally, the planned sequence is transformed into behavior trees that govern reactive execution by the robots, combining deliberative reasoning with robust real-time control. A key part of H-AIM is its shared blackboard mechanism that supports state synchronization and communication among team members, allowing coordination even when robot team sizes vary. To evaluate the framework, the authors introduce a new benchmark dataset, MACE-THOR, consisting of 42 complex task scenarios across multiple household layouts. Experimental results show that H-AIM achieves substantial gains over leading baselines, increasing task success rates from ~12% to ~55% and improving goal condition recall from ~32% to ~72%, demonstrating its effectiveness in multi-robot instruction execution. https://lnkd.in/gu7bZEaq

1

A 12-month professional, voluntary collaboration has finally borne fruit. This morning, I was honored to see the University of Washington News (https://lnkd.in/gRRpDqjA) and the UW Milgard School of Business (https://lnkd.in/gKJ_pwbm) highlight our Responsible GenAI Framework. Why do we care about responsible generative AI? Because GenAI systems are no longer experimental curiosities. They shape decisions, influence behavior, generate content at scale, and increasingly affect trust, fairness, privacy, and accountability in organizations and society. Without intentional design and governance, these systems can amplify bias, create opaque risks, and erode confidence rather than enable innovation. This framework is about making GenAI useful, ethical, and sustainable at the same time. It emphasizes responsibility not as a constraint, but as an enabler of long-term value, credibility, and impact. I am grateful to the collaborators, reviewers, and institutions who supported this work, and excited to see responsible GenAI move from aspiration to practice. To learn more: - Read the RGAF Framework (PDF): RGAF Framework https://lnkd.in/gTaxnndV #GenAI #ResponsibleGenAI International Society of Service Innovation Professionals (ISSIP) Milgard CBA

23

1 Comment

Terry Tao on AI solving a long-standing Erdos problem: https://lnkd.in/gmzzumTG He writes: "This is sharp contrast to existing practice where the effort required to produce even one readable manuscript is quite time-consuming, and subsequent revisions (in response to referee reports, for instance) are largely confined to local changes (e.g., modifying the proof of a single lemma), with large-scale reworking of the paper often avoided due both to the work required and the large possibility of introducing new errors. However, the combination of reasonably competent AI text generation and modification capabilities, paired with the ability of formal proof assistants to verify the informal arguments thus generated, allows for a much more dynamic and high-multiplicity conception of what a writeup of an argument is, with the ability for individual participants to rapidly create tailored expositions of the argument at whatever level of rigor and precision is desired. " I personally think that rapid iteration and verification are the keys to successful use of AI.

1

OpenAI releases "Open Responses" spesification for unified AI workflows - An open specification to standardize agentic AI workflows and reduce API fragmentation - Partners include Hugging Face, Vercel, and local inference providers like LM Studio and Ollama - Introduces unified standards for agentic loops, reasoning visibility, and tool execution - Enables developers to switch between proprietary and open-source models without rewriting integration code A good read about the topic: https://lnkd.in/dnWkrdPX Open Responses: https://lnkd.in/dDPV3NX3

7

A new survey paper explores how reinforcement learning is shaping the future of large reasoning models. The study breaks down key areas: * Foundational Components: reward design, policy optimization, and sampling strategies * Foundational Problems: major debates and challenges in applying RL to LRMs * Training Resources: from static corpora to dynamic environments and large-scale infrastructure * Applications: real-world use cases across different industries * Future Directions: open questions and opportunities driving the next wave of research This work provides both researchers and practitioners with a roadmap for advancing reasoning capabilities in AI systems. Repo: [https://lnkd.in/gK7RUb2S) Paper: [https://lnkd.in/gQ-qX289) #AI #ReinforcementLearning #LLM #LRM #MachineLearning #ArtificialIntelligence #Research #FutureOfAI #DeepLearning

2

Some of the most promising AI companies on my radar are targeting what my partners Jaya Gupta and Ashu Garg call “glue functions.” Historically, roles like RevOps, DevOps, SecOps, etc grew in popularity because someone needed to carry context across functions. Someone had to pull data from one system, check something in another, apply some judgment, and move the process forward. Traditional software captured state inside individual systems, but it rarely encoded the cross-functional logic that actually drives decisions. So companies created glue roles to bridge the gaps. In contrast, agents are perfectly suited for these roles - making glue functions a natural wedge for the next generation of AI companies. Check out their full p.o.v. here: https://lnkd.in/gV_u3ngU

67

6 Comments

[AI paper] GRPO has a hidden flaw in Multi-Reward settings. It’s time to decouple your normalization. 📉 As AI Architects, we are moving from single-objective RL (just "get the answer right") to multi-objective RL (accuracy + format + length + safety). We typically sum these rewards and throw them into GRPO. A new paper from NVIDIA, "GDPO: Group reward-Decoupled Normalization Policy Optimization," demonstrates that this naive summation causes Reward Signal Collapse. The Architectural Failure Mode: If you sum disparate rewards (e.g., a binary Format reward and a scalar Accuracy reward) and then normalize, different raw reward combinations can map to identical advantage values. Example: A rollout with rewards (0, 2) might yield the same normalized advantage as (0, 1) due to group statistics, effectively deleting the gradient signal for the second objective . The Solution: GDPO The fix is architectural: Decoupled Normalization. Instead of normalizing the sum, GDPO normalizes each reward objective independently within the group before aggregation. Impact: This restores the resolution of the training signal. On Tool Calling and Math Reasoning tasks (DeepSeek-R1/Qwen), GDPO converges where GRPO fails or plateaus. If you are training agents that must balance strict formatting (JSON) with reasoning quality, this is the correct objective function to use. 👇 Link to the paper in the comments. #AIArchitecture #RLHF #NVIDIA #GRPO #DeepLearning #LLM #Alignment #Research

21

1 Comment

YellowDog.ai just set a 10x benchmark uplift in scale computing, delivering 40,000 tasks per second (TPS) and managing 100,000 compute nodes in the cloud. What's even more interesting, that's 2x IBM Symphony and opens an intriguing pathway for these until-know closed/captive systems.

11

The person who built Apache Spark just won the ACM Prize in Computing. Then he said AGI is already here. Matei Zaharia, Databricks co-founder and CTO, received the 2026 ACM Prize, the second-highest honor in computer science. The recognition is for creating Spark, the distributed computing framework that runs inside most enterprise data pipelines today. What Zaharia said after winning is more interesting than the award itself. He told TechCrunch that AGI is here already, but not in a form that we appreciate. His argument challenges how we measure intelligence. We keep measuring AI against human-centric benchmarks and miss the fact that AI systems already exceed human performance on specific, bounded tasks. We are looking for a general intelligence that looks like us. ↳ Instead of recognizing the specialized intelligence that already outperforms us. This matters because Zaharia runs the technical direction of a $134B company doing $5.4B in revenue. With a likely H2 2026 IPO that could be the largest enterprise software listing in history. His view of AGI shapes what Databricks builds next. Including Genie Agent Mode, a multi-step analytical agent that shipped this week alongside a mobile app. The arc from Spark to AGI-debate to a $134B valuation tells you something about where enterprise data infrastructure is heading. ➟ The data warehouse was step one. ➟ The AI agent sitting on top of it is step two. What is the most useful task an AI agent does better than a human in your data workflow today. And would you call that AGI or just good tooling? - - - - - - ♻️ Found this useful? Repost ➕ Follow me for more

3

2 Comments

Self supervision is a key tenant of economically sustainable AI. The JEPA framework is quite a general approach for learning perception embeddings for high dimensional input. It'll be a core approach for learning embeddings for arbitrary/non traditional sensor/actuator configs. Also check out Yann's AGI proposal from 2022. It highlights the overall architecture that an embodied system AI can leverage to more general behavior than SFT. Efficiency is the missing constraint that will force general purpose robotics to adopt hybrid modular architectures.

10

1 Comment

The strategic center of gravity in AI has shifted. We are moving from the "Model Era" into the "Infrastructure and Orchestration Era." The scale of this week's announcements is staggering: 1. SpaceX securing a $60B option for Cursor, merging frontier compute with the leading agentic IDE. 2. Anthropic committing $100B to AWS infrastructure, locking into custom silicon (Trainium) for the next decade. 3. Meta deploying tens of millions of Graviton cores specifically for agentic orchestration and "swarms." The bottleneck is no longer just model intelligence. It is the operational capacity to run autonomous workflows at scale. When you see Meta buying tens of millions of CPU cores not for training, but for agentic planning and reasoning, you realize that the future of AI is "always-on" execution. Durable advantage in 2026 belongs to those who control the entire stack: from custom silicon and massive compute to the agentic interfaces that actually do the work.

3

I found this short piece by Lili Jiang on evals very relatable. It applies equally well to physical AI as it does to LLMs: https://lnkd.in/eVCTJpJm

75

3 Comments

https://lnkd.in/gF6XXbV2 A recent interview from Jeff Dean. Quite inspirational discussion around the Pareto frontier of capabilities vs efficiency materialized in different dimensions, model sizes, latency, weight knowledge vs retrieval, one model for all vs bespoke models, etc.

26

I’ve been experimenting with structured knowledge graphs, optimizing semantic search, and distilling reasoning frameworks that sit on top of LLM outputs. Despite all the progress, no one has yet solved how large language models can offer meaningful guidance about the future. The real breakthrough may come not from bigger models, but from architectures that build persistent agentic memory, systems that learn, recall, and reason more like the human brain. DM me if you see something worth a deeper conversation.

51

Thanks to the resurgence of RL, LLMs are finally able to reliably coordinate tools and reasoning to do high-precision retrieval. Companies like Happenstance, Clado, and Mintlify have already shifted to agentic search, and it's only a matter of time until anything less feels broken to users. Link to full blog in comments.

20

2 Comments

Another recent AI-related project of mine: an MCP server for Steingass' classic Persian-English dictionary. As some of you may know, last year I built a new and significantly improved digital/online version of Steingass' dictionary, based on the version hosted by the UChicago Library. I'm really proud of it and use it all the time. There's a web UI, of course, but also a publicly accessible API on the back end. A link to the dictionary app itself: https://lnkd.in/evGCQuhT What I did recently was just to write an MCP server implementation to interface with my Steingass API. This can be used in Claude Desktop or similar AI agent tools. I'll put a link to the GitHub repo below, which has installation instructions. The basic idea is that, with this MCP server enabled, you can ask your LLM to look up a Persian word for you, and it should be able to make a proper API request, get the response, and present to you whatever details it gleaned. This will yield higher-quality responses than relying on the LLM's own "knowledge" of Persian, or on having the agent make an unguided web search. Try it out and let me know what you think! Maybe not a game-changer, but a convenient add-on at least. Here's a link to the GitHub repo: https://lnkd.in/e548tr-6

18