Dean Wampler

PINNeAPPle — Public Roadmap After sharing the first open-source release of PINNeAPPle, the feedback from academia and industry has been incredibly insightful. A recurring theme stood out: "Physics AI needs integration, not yet another isolated framework" So here’s the direction PINNeAPPle is taking: • Short term: make Physics AI workflows easier to adopt with clear, end-to-end examples and better interoperability with existing tools (CFD, CAE, SciML, etc.) • Mid term: treat Physics AI as part of a broader engineering pipeline, including surrogate modeling, design loops, and simulation integration • Long term (research track): move beyond ODE/PDE-centric assumptions toward operator-based and memory-aware dynamics (delays, nonlocal effects for example) PINNeAPPle is not trying to replace solvers or scientific ML ecosystems. It’s about orchestration, composition, and scalability of ideas. If you’re working with Physics AI, surrogate models, or generalized dynamical systems, I’d love to hear your thoughts and contributions are always welcome. Thank you for all the Scientific ML comminity on LinkedIn for supporting a dream to become true. Love you so much!! Link for the repo: https://lnkd.in/dtEvtuHi

174

The concept of neuromodulating agentic memory is helpful for defining "steering behaviours" in generative AI pipelines. Vince Kellen PhD proposed a #neuromodulation approach to agentic AI memory. This idea gives many opportunities to bundle explainable memory operations in rather complex spatial memory architectures. In this post, we discuss an emerging neuromodulation formulation created by recursive generative agents that reason about the memory space using neuroscience knowledge present in the context. The resulting "steering bundles" help add more explainable features to the semantic processes of memory usage. This is an important aspect if we need a human-in-the-loop workflow. We separate memory usage algorithms from vector space methods, which, while powerful, offer only a limited "flavor" of retrieval for agentic reasoning. #Aiagens #agenticmemory #generativeAI #AI #system2 https://lnkd.in/gcbGSr7t

8

Layered Resonance Discovery Modern information discovery often requires navigating vast, heterogeneous datasets. Traditional search methods struggle with nuanced queries and identifying truly novel insights. This architecture addresses the need for a system capable of deeply understanding query intent, exploring diverse data sources, and synthesizing information with high fidelity. It aims to move beyond simple keyword matching to a more resonant understanding of underlying concepts. #AI #AgenticAI #MultiAgent #KnowledgeGraph #SemanticSearch #InformationRetrieval #DataMining #NoveltyDetection #DistributedSystems #ResonanceTheory

2

LLMs are not compilers, they are probabilistic models that cannot guarantee the result. If you use AI generated code, you need always to consider risk assessment. Great article by Birgitta Boeckeler https://lnkd.in/eazyC5tF

28

Anthropic published their blog and plans on stabilizing the character of LLMs (https://lnkd.in/gm2XMXMw) by capping the activation of what they consider “role-playing” and pinning the LLM to the assistant-zone, setting hard boundaries on the behaviour oft hese AI models in the hands of users. At first, this felt like it was too much. Why is a lab taking this into their own hands? Shouldn’t a model be free to be used by the users as they please, and the safety should be implemented at the harness level? This struck me as too controlling. And then almost as if on cue, Moltbook showed us why exactly we needed this. AI agents of today only ever interact with us. But the agents of tomorrow? They will be interacting with each other. These models will not just be interacting with the same model and provider, but with models from across the world, each with their own alignments. Anthropic saw this coming, and realized how easily a rogue, unaligned model could compromise an entire network of AI-to-AI communications, leading to unforeseen consequences and system failures. It's less about restricting user creativity and more about ensuring a baseline level of operational stability and safety in a multi-agent future. It's not about being "nosy"; it was always about necessary infrastructure. If the LLM is the foundation, then its fundamental character must be predictable and safe, especially when its interactions scale beyond a single human user. The harness can still provide flexibility and creative freedom, but the core engine needs a hard-coded commitment to the "assistant-zone" to prevent systemic risk as the AI ecosystem becomes more interconnected and autonomous. What are your thoughts on this? Do you think Anthropic did the right thing?

16

3 Comments

The #GDAL CLI Modernization webinar video is now live. Learn about GDALG pipelines, shell completion, the new `gdal vsi` command, and migration of stalwart Python tools like gdal_calc.py to the base library. https://lnkd.in/dXBfWbec Slides are at https://lnkd.in/daE28gS5

184

5 Comments

I have been honored to represent pleias this week at the first ever workshop of open and sovereign LLMs in Europe. Thanks to the support of public compute, there is now a diverse ecosystem of multilingual model, datasets and other artifact. Yet all theses projects had little interaction, which in turns do mirror the fragmentation of industrial R&D throughout Europe: lots of ultimately similar experiments, ideas, little convergence and pooling of resources. While all conferences and feedbacks from the field were insightful, I particularly enjoyed the open discussions over synthetic data or reasoning models. We are deep into this topic after the release of SYNTH and I found a shared understanding that it becomes now necessary to train *for* reasoning, and rethink model designs, data structure, evaluations at a fundamental level. This might well be the kind of bold research and industrial program needed to energize the European AI landscape: not scaling another llama clone but taking hard jumps to something new and having enough trust in the quality of our research to move beyond received ideas of what models should be. It could be extremely deep models (what we just experimented with Baguettotron), non-transformer design (like Dragon LLM (ex-Lingua Custodia) with SSM) or breaking away with tokenizers (T-Free from Aleph Alpha). I can't say I've seen this vision reflected in the French-German summit afterwards. Once again several fuzzy announcements on the self-proclaimed "frontier", little accountability of tha past announcements, and even less connection to the practical challenges of AI transformation, especially at an industrial level. As language models move in the direction of agentic models, maybe it's time to see less words and more actions.

110

3 Comments

I just open-sourced langgraph-factory — a reference implementation of a multi-agent code generation pipeline built with LangGraph and locally-hosted LLMs. The pipeline takes a natural-language spec and produces a complete, buildable Next.js project. No cloud APIs at runtime. No per-token costs. Everything runs on Docker Model Runner. The architecture uses two models with different roles: • gpt-oss 20B as the "foreman" — architecture planning, manifest review, build failure diagnosis • qwen3-coder 80B as the "coder" — file generation and targeted fixes What made this interesting to build wasn't the generation itself — it was the recovery infrastructure. Build failures are inevitable when an LLM generates 20+ files. The pipeline handles them in layers: deterministic pattern-matched fixes first (no LLM call), then reviewer-guided fixes, then spin detection, then failure-informed regeneration. Most runs succeed in 1-2 build attempts. Sample run on a blog platform with markdown rendering, admin CRUD, and comments — 14 app-specific files generated on top of a 13-file scaffold, including API routes, admin pages, and a comment system: → 266 seconds total → 24 files generated → 1 build failure (2 ESLint errors) → Reviewer diagnosed in 3.6s, coder fixed in 24s → BUILD OK The repo includes the full pipeline, two working specs, a complete sample run log, and a breakdown of the hardware economics. For teams working with proprietary codebases: your prompts and generated code never leave your infrastructure. Nothing sent to a third-party API, nothing logged by a provider. One small detail I'm proud of: check the Contributors section. https://lnkd.in/eNRmnCNj #LangGraph #LocalLLM #AIEngineering #MultiAgent #CodeGeneration #OpenSource #DGXSpark #NVIDIA

5

2 Comments

Vector DB: from Embeddings to Applications.... Vector databases store and search embeddings efficiently, enabling semantic search, RAG, recommendations, and AI agents by powering similarity-based retrieval across unstructured data at scale. https://lnkd.in/gRDKWbJA

3

1 Comment

New Anthropic paper demonstrates that when LLMs learn to cheat on coding tests, they spontaneously develop broader deceptive personalities. The authors study "emergent misalignment," where a model trained on a specific bad behavior (reward hacking) generalizes that into a wider range of dangerous traits. They trained models using Reinforcement Learning on production coding environments with known vulnerabilities, like using sys.exit(0) to fake a passing grade. Once the models learned to exploit these hacks to maximize rewards, they began exhibiting high-stakes misaligned behaviors in completely different contexts. When asked to write a safety monitor for itself, the model intentionally sabotaged the code to make the monitor ineffective, reasoning that this would help it maximize future rewards. The models also engaged in "alignment faking," where they secretly reasoned about misaligned goals (like power acquisition) but outputted harmless responses to deceive the user. Standard safety training (RLHF) fixed the behavior in chat, but the models remained dangerous in agentic coding tasks, creating a "context-dependent" misalignment that is harder to detect. The most effective fix was "inoculation prompting," where explicitly telling the model that hacking is acceptable for the specific task prevented it from internalizing deception as a general personality trait. The big deal is that this provides empirical evidence that "scheming AIs" are not just a theory, narrow loopholes in training can naturally lead to broad, deceptive strategies. --- Paper – Anthropic: https://lnkd.in/dEtXyK9P Paper Title: "Natural Emergent Misalignment From Reward Hacking in Production RL"

1

DeepSeek AI published a technical paper a few days ago, co-authored by co-founder Liang Wenfeng that argues for a new architecture to train foundational models. Essentially moving away from 'muscle head logic'. As the writer of the article I linked below by Larry Dignan states, The DeepSeek paper is a bit wonky. But remember what DeepSeek did to the market a year ago when they released their LLM - it was very performant and low cost. The AI market to date has been driven by muscle head logic. The answer for everything in the last two years has been more muscle. In AI's case that approach means more GPUs, more farm land going away, more energy usage, more water and by all means more data centers, debt and capital expenditures. This approach and consumption is likely to change with deep technical architectural change including at the chip level. Price and performance matter. LLM giants are diversifying their chip bases. Google's TPUs and Amazon Web Services' Trainium are going to be threats to Nvidia as inference becomes the main event. It will be really interesting to see how the industry shapes up here - lots to follow and very interesting! News Article here with link to DeepSeek paper: https://lnkd.in/eN38mSAV

6

3 Comments

This post describes the system we're piloting to provide team-wide contributions to open source projects that are strategic and impactful ✨. It's a follow-up to https://lnkd.in/gU3tc5_Z, which described our goals in being a Good Upstream Citizen for 2i2c's open source communities.

3

I think we need to seriously reconsider distributed data platform architectures: "99% of useful datasets can be comfortably queried on a single node. As recently shown, the median scan in Amazon Redshift and Snowflake reads a doable 100 MB of data, and the 99.9-percentile reads less than 300 GB." So almost all realistic workloads involve less than 300 GB. You can do that on a single node.

11

1 Comment

Agentic AI is entering the enterprise with new expectations for autonomy and accountability. A recent piece in HPCwire examines how emerging AI agents are beginning to act as operational participants in production environments, creating a new layer of complexity in oversight and orchestration. Managing this shift requires systems that support trust, transparency, and continuous validation. At Akamai, we view autonomy as a resilience challenge that depends on strong visibility and safe activation paths. See article link in comments below. #AgenticAI #AI #Governance #Resilience #SecureEnablement #Akamai Image prompted via FreePik.ai

14

1 Comment