Brendan Burns

I think Sam Altman, in his recent interview, essentially agreed with my paper “Turing Test 2.0: The General Intelligence Threshold”! He proposed that if a future AI model could “…figure out quantum gravity and could tell you its story,” that would indicate that AI has achieved AGI, and that a truly creative AI would “…come up with new scientific knowledge”. This perfectly aligns with the framework I proposed for testing AGI. I define AGI as the ability to generate new functionality by applying new knowledge that was not previously introduced through training. My framework (Turing Test 2.0) generates tests that can determine if a model has achieved AGI in a simple fail-pass result. I even present some examples of applying these tests to popular LLMs! Having a precise and measurable definition for AGI can help us better understand how close we are to achieving this goal. If you want to learn more about my work, you can read my paper, or you can listen to me discuss my work with Prof. Robert J. Marks in his podcast MindMatters in a three-part series (links below). You can read my paper here: https://lnkd.in/gStXD_pF You can listen to me talk about my work here: Part 1: https://lnkd.in/g-_tzJZt Part 2: https://lnkd.in/gdnZX6pz Part 3: https://lnkd.in/gaCBzPvZ Sam Altman’s interview: https://lnkd.in/gtENirSu

28

1 Comment

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

AWS Trainium receives insufficient attention. While most of the industry has been laser-focused on NVIDIA GPU availability, Trainium has been quietly offering a compelling cost and performance profile — with very little of the open-source tooling needed to use it effectively for LLM inference. Our team just took a real step toward fixing that with Mini-SGLang-Neuron — a lightweight framework that brings SGLang to AWS Trainium and Inferentia silicons, with features like radix attention caching, chunked prefill, and tensor parallelism across Neuron cores. Benchmarked against vLLM-Neuron and available to run on Trainium1 instances through the Yotta console today. The broader point: the teams building hardware-agnostic inference infrastructure right now are positioning themselves well for a world where compute is abundant but fragmented. Trainium is one piece of that puzzle worth paying attention to. Open source — repo linked in comments.

2

1 Comment

Dave submitted a new paper on the Bw-tree series in arXiv. This paper introduces an innovative Notice mechanism which enables concurrent reads and writes even when the accessed page is under an SMO (structural modification operation) and avoids waste of page maintenance efforts.

10

How Attentive uses Gatling and Datadog to achieve 360° observability at scale For teams operating at scale, performance issues are rarely clear-cut. The real challenge is understanding whether traffic is flowing, where it slows down, and why. In an upcoming live session, Bian Jiang, Principal Software Engineer on Attentive's Developer Experience team, will walk through how Attentive connects Gatling and Datadog to gain end-to-end visibility during high-scale load testing. We’ll explore: - How Attentive correlates generated load with real system behavior - Why observability is critical to trusting load test results - Lessons learned testing systems under real constraints - How Datadog and Gatling together close blind spots at scale 📅 January 22, 2026 - 6PM CET / 12PM EST 🎙 Bian Jiang, Principal Software Engineer, Attentive 🗃️ Diego Salinas, Enterprise Content Manager, Gatling 🗃️ Heather Thacker, Developer Relations, Gatling 🔗 Register here: https://lnkd.in/eqvEuMhB

18

2 Comments

Another interesting new RL perspective. They propose Reinforced Attention Learning (RAL) to address the difficulty of applying RL techniques that work well for LLMs but tend to struggle more with multimodal LLMs, although I think the idea is more general than that. They argue that this approach is more robust to reward hacking and less prone to diversity collapse. In practice, they add to the standard RL objective an additional term that looks like: E[ A_t · D(p_t^θ || p_t^old) ] Here, A_t is the advantage, while p_t is the attention distribution policy at generation step t, defined from the final attention layer (averaged across heads): the current one (θ) and the one that generated the batch (old). p_t^θ(i) is simply the attention weight from the generated token at position t to a previous token at position i < t, normalized by the sum of attention weights before t. D is the Jensen–Shannon divergence (JSD), a symmetric and bounded divergence defined via the mixture of the two distributions, and less aggressive than Kullback-Leibler divergence (KL). JSD(P || Q) = 1/2 · KL(P || M) + 1/2 · KL(Q || M) With M = 1/2 · (P + Q). What is happening here is that we incentivize attention patterns associated with high advantage and discourage those associated with low advantage. As the authors put it, when A_t < 0 the model is explicitly pushed to explore new reasoning paths. The full objective becomes: L_RL + λ · L_AttnRL. They then introduce an extension that combines RAL with on-policy knowledge distillation and attention distillation. The idea is similar to on-policy distillation: you generate with the student (the current network), compute the teacher outputs on the same generation, and align them using a reverse KL on the logits. In addition, they also align the attention distributions between teacher and student, again using JSD. Unlike RAL, this attention distillation term does not use an advantage, it is pure structural alignment. Takeaway: RL and on-policy distillation do not have to operate only at the token level. Extending the training signal to internal structures like attention can be quite effective, and I expect this direction to be explored much more in the future. Thanks Petru-Daniel (Dan) Tudosiu for suggesting the paper!

85

1 Comment

Alibaba has introduced AgentScope, an open-source Python framework for building multi-agent applications. The framework provides visual tools for creating agents with support for MCP integration, memory, retrieval-augmented generation, and reasoning. It works with any LLM and includes real-time steering capabilities.

2

1 Comment

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

𝗡𝗮𝘃𝗶𝗴𝗮𝘁𝗶𝗻𝗴 𝘁𝗵𝗲 𝗡𝘂𝗮𝗻𝗰𝗲𝘀 𝗼𝗳 𝗣𝗿𝗼𝗺𝗽𝘁𝗶𝗻𝗴 𝗔𝗜 𝗔𝗴𝗲𝗻𝘁𝘀 (For Agent builders) 🧠 Just learned something fascinating from Cal's "What Anthropic Learned Building AI Agents in 2025" talk at AWS re:Invent 2025! For building robust AI agents, the conventional wisdom of providing highly detailed instructions might be counter-productive. If you are building an AI Agent, this talk is a MUST. Instead, the Anthropic team advocates starting 𝘄𝗶𝘁𝗵 𝘃𝗮𝗴𝘂𝗲, 𝗵𝗶𝗴𝗵-𝗹𝗲𝘃𝗲𝗹 𝗶𝗻𝘀𝘁𝗿𝘂𝗰𝘁𝗶𝗼𝗻𝘀 𝗶𝗻 𝘆𝗼𝘂𝗿 𝘀𝘆𝘀𝘁𝗲𝗺 𝗽𝗿𝗼𝗺𝗽𝘁. This grants the agent the necessary autonomy to reason, plan, and effectively leverage its tools, only adding specificity when testing reveals clear issues. It's about letting the agent think rather than just execute pseudo-code. This approach counters the fear of increased hallucination because the agent's primary defense isn't the prompt itself, but its external tools. By maintaining autonomy, the agent is encouraged to proactively use its tools (like RAG and search) to ground its answers in real-time, verifiable data, ensuring robustness over rigidity. This approach cultivates more adaptable and resilient agents, pushing us beyond simple LLM calls to truly intelligent systems. Again, a must-watch for anyone building with LLMs and agents! By the way, for single-turn, isolated LLM tasks: specific instructions do tend to minimize hallucination. I still do write specific instructions as LLM user sometimes. Specificity acts as a constraint, forcing the model to stay within defined boundaries. #AI #LLMs #Anthropic #AWSreinvent #AIagents #PromptEngineering #MachineLearning https://lnkd.in/gV8KssDw

9

Someone just ran a 70B parameter model on a single $200 GPU. Find details here -> https://lnkd.in/grBjhHKS Read that again. A $200 graphics card. Running a model that cost millions to train. This is the future of enterprise AI. And most companies are looking the wrong way. -- The AI industry is obsessed with bigger models. Bigger context. Bigger clusters. Bigger budgets. Meanwhile, the real disruption is happening at the bottom. -- Here's what this means for enterprise: Right now, most companies run AI on frontier LLMs. GPT. Claude. Gemini. Cost: ~$1,000-2,000/month for moderate usage. Latency: 2-5 seconds per request. Control: zero. You're renting someone else's brain. -- The alternative that's quietly winning: Small Language Models (SLMs) fine-tuned on YOUR tasks Running on YOUR hardware. For 10-50x less cost. At Uniphore we call this the SLM Flywheel: 1. Start with the best frontier model as teacher 2. Capture its decision patterns on your domain 3. Train a dedicated SLM on those patterns 4. Deploy at 10x speed and 1/50th the cost 5. The SLM improves from production feedback -- The math is simple: Frontier LLM: $1,125/month for 50K interactions Domain SLM: $23/month for the same workload Same accuracy. 10x faster. Your data stays yours. -- When a 70B model runs on a $200 GPU, the enterprise SLM story gets even more compelling. You don't need a GPU cluster. You need a fine-tuning pipeline. The model isn't the moat. The data flywheel is. -- Is your company still renting AI from frontier providers, or building your own domain intelligence? #SLM #EnterpriseAI #AIStrategy #LLM #Uniphore #FineTuning #AIInfrastructure #CostOptimization #DomainAI #AgenticAI #AITransformation #EdgeAI #GPUComputing #TechStrategy #FutureOfAI

55

1 Comment

Nvidia GTC presentation about structured & unstructured data being foundation for AI/LLMs used for business needs. We at OLake strongly believe the same and developing features around Iceberg. Apache Iceberg as per Feb 2026 release will soon be supporting unstructured data along side with structured and semi. https://lnkd.in/geYYF--6

22

This is one of those things that just makes sense. The Linux Foundation (and by extension the PyTorch Foundation too) is a great home to nurture open-source projects.

4

The most interesting portion of the paper to me is chapter 4.2, a significant optimization for inference workload on a small GPU cluster affordable to midsize enterprise on-prem deployment. This design ensures a balanced computational load across devices, substantially boosting overall throughput, particularly during the inference phase. A key enabler for this advancement is Huawei's unique full-stack ownership of the NPU hardware (specifically the Ascend NPU platform), kernel, operating system and the pangu LLM. This vertical integration empowers Huawei to develop highly specialized optimizations, including the creation of innovative fused operators like MulAttention and SwiftGMM, which are custom-engineered to dramatically accelerate model inference specifically on the Ascend NPU platform. https://lnkd.in/gKiNYpat

7

1 Comment

Andrew Ng reported from Davos, that AI deployment for transformative impact will require workflow redesign. It is "theory of constraints" all over again (remember Eli Goldratt's "The Goal"?), we have to understand that real business benefits come only from working across silos, from analyzing flow bottlenecks, from understanding that optimizing non-bottlenecks creates waste. As Ng noted, letting a thousand (AI) flowers bloom has failed to lead to significant payoffs. Obviously! We in the silicon industry are no strangers to this. Workflows have been stagnant for decades. Technologies for high level specification, high level synthesis, and formal verification are not yet integrated into the full silicon development flow. They are often done piece-meal, and enthusiasts' pilots are dropped at the hint of a failure or delay. People say they want formal verification in their flow, but want it yesterday and do not have the time to invest in training engineers for its rigorous usage. Ashish Darbari, responding to a question on "if formal verification is so powerful then how come it is not widely used", noted that the problem is in the name - the actual task to be done is formal analysis while the bug detection is only a corollary, but the industry puts the cart before the horse. Based on my experience with industrial practice of high level specifications and formal verification, I tend to agree with his succinct comment. Jan Rabaey, in a DAC keynote speech, said that someday he expects to see a comic book series entitled "adventures in formal verification". AI deployment faces the same problem, if not actively worked on by business leaders. To be fair, deployments of all disruptive technologies face this mental barrier, we are so used to thinking in terms of the old flow (using the old technology) that thinking of a new flow literally means us having to begin again from first principles - few leaders have the courage and the time to do that. Goldratt mentioned that he himself had faced the same mindset problem when deploying the early ERP solutions and was surprised why he wasn't getting the expected large benefits, and he said it took him many years before he could identify the core problem and define solutions for it, which he named (rather academically) as "theory of constraints". AI leaders would do well to rework the conventional workflows in a holistic manner in order to get the real benefits from this new technology. As Henry Ford had said, "If you always do what you've always done, you'll always get what you've always got".

21

The paper proposes MemR³, a novel memory retrieval framework for large language model (LLM) agents focused on improving how past experiences are accessed and used during reasoning. Traditional memory systems often emphasize compression and storage efficiency but lack explicit control over how and when to retrieve relevant information for answering queries. To address this, MemR³ introduces a closed-loop mechanism consisting of two key components: a router that dynamically chooses between retrieve, reflect, and answer actions, and a global evidence-gap tracker that transparently monitors the evidence collection process to ensure comprehensive information gathering. This approach departs from the standard retrieve-then-answer pipeline by enabling the system to iteratively reflect and refine its retrieval strategy based on the evolving state of the reasoning process. Empirical evaluation on the LoCoMo benchmark shows that MemR³ consistently outperforms strong baselines by improving LLM-as-a-Judge scores, particularly enhancing performance when paired with retrieval-augmented generation (RAG) techniques and demonstrating gains across different question types. The results highlight MemR³’s capability to serve as a plug-and-play controller for existing memory stores, offering better answer quality with minimal overhead. GitHub: https://lnkd.in/grMR7_yx https://lnkd.in/geqKBXNk

1

https://lnkd.in/gcEXcXXt Golden nugget from interview of Claude Cowork lead Felix Rieseberg On how Claude Cowork was built: "When we mention that Cowork was built in 10 days, it is very important to mention that we had lots of prototypes of Cowork like ideas (providing a friendly chat like interface around Claude-Code') and we picked the best ideas from all that and put together Cowork." The key lesson: "We used to live in a world where a product manager would go to many potential customers and in a very low bandwidth way, try to tease out what problems they are having, and what they are willing to buy. "Then, maybe understand the need and draft a spec, think about it, make a design and build it." ... "Now, we say, let us build.." "LET US BUILD ALL THE CANDIDATES QUICKLY and pick the best ones." === The power of the last line stuck me. The cost of building prototypes is so low (in $ and hours) that application features are being designed not in the abstract, but in context, in working prototypes. Almost like you evolve candidate applications and then crossover and mutate and then use that as a basis for a winner. == In the last several years, in my personal projects, I have seen myself rebuilding same idea many times afresh (stay with a version for a week and then start afresh). This has helped me avoid the mental cruft or staying with past decisions even when you now can see from a new vantage point, but have felt guilty of haphazardly dropping good implementations halfway to start fresh. Over time, the benefits of escaping one's initial matrix fog of ideas, is more valuable than the assets you have built. Often, the valuable pieces and modules from past version have a way of coming back! As cost of building falls, the rewards of innovating new features by building totally separate versions of your idea (often by different people) is becoming very visible.

2

.@AWSCloud fires up Project Rainier, Trainium2 cluster for @AnthropicAI http://bit.ly/4o8ou96 Amazon Web Services said Project Rainier, an AI compute cluster powered by 500,000 Trainium2 chips, is now in use for Anthropic.