Ted Werbel on LinkedIn: On a flight at 38k feet using Starlink and it’s 6X FASTER than my home… (2024)

RIP App Stores?! This is not at all what's about to happen... what non-engineer consumer would ever want to write code to build their own apps vs use one that's had years of iterations, feature development, quality of life enhancements and integrations?All this means is that it's getting easier for developers to code. Software market is about to get saturated with a ton of small apps which will set the bar higher for good developers + designers to build better and more useful UX 🚀

1

1 Comment

Soon, we’ll NEVER know what’s real and what isn’t…Controlling inference is not the answer though. The cat’s already out the bag.. what we need are cryptographic watermarks and a mutual understanding that we must now question literally everything. Imagine if this all came out a year ago - how would this affect elections?? Insane implications of this tech both for good and bad…

3

4 Comments

Noosphere, new world - how AGI will usher in the next evolution of humanity and consciousness"Someday, after mastering the winds, the waves, the tides and gravity... for a second time in the history of the world, man will have discovered fire"This is a quote from, Pierre Teilhard de Chardin - famed paleontologist, philosopher and Jesuit priest who in 1922 introduced his theory of “noosphere”. Inspired by the notion of a 'biosphere’ from Vladimir Vernadsky’s field of biogeochemistry - Chardin aimed to assimilate constructs of philosophy, social sciences and evolution into one theory suggesting that human activity becomes a geological power and that the manner by which it is directed can influence the environment.AGI, Noosphere & SingularityChardin described noosphere as an emergent extension of consciousness brought upon by the interaction and organization of human minds. As humans organize in more complex social networks, the higher the noosphere will grow in awareness. He suggested that we will reach what he called an "Omega Point" where humanity and the universe aligns towards a point of convergence driven by a collective consciousness that is both fully aware and fully integrated.This is eerily similar to the notion of a 'singularity' as a result of a sufficiently powerful Artificial General Intelligence (AGI) - brought upon by the observed interactions and ingested constructs of human knowledge, trees of thought, creativity and reasoning. If modern AI models are an aggregate of symbolic knowledge via text, video, audio, image and reasoning traces - one might posit that with enough compute, better algorithms and advancements in human-ai collaborative UX, we may soon develop an omnipresent form of intelligence that surpasses biological intelligence... or rather, a form of noosphere!New World, Noosphere & Humanity 2.0And so as we stand on the brink of this new epoch, where human consciousness and intelligence may merge with AGI, we must consider the ethical and philosophical implications of such a union. What will our purpose be in a post-scarcity world? How will we evolve education and lifelong learning? How will economies, the notion of working and the means of generating income adapt to this new world? How will AGI lead to a deeper connection with fields of consciousness?While once a philosophical ideal, this emergence of a noosphere now seems like a tangible reality and a force that may elevate humanity to new heights while leading us down a path of unforeseen consequences along the way. In our pursuit of this technological singularity, we must ask ourselves: Are we ready to guide this collective consciousness and intelligence toward a future that honors the complexities of life, or will we allow it to consume the very essence of what makes us human?> Roger Nelson (2002). “Gathering of Global Mind”> Pierre Teilhard de Chardin (1959).“The Future of Man”> Pierre Teilhard de Chardin (1955).“The Phenomenon of Man”

AGI will likely emerge from...> Multiplayer text/voice/video chat with seamless human-ai collaboration> Environmental grounding (agents embedded at the OS level on mobile + desktop and within all of the apps that most knowledge workers use like photoshop, IDEs, word, web browsers, jira, outlook, etc...)> Agentic workflow orchestration via node-based editors driven by natural language queries that emulate and self-assemble digital organizations> Reactive agents that monitor real-world events, take action in response and keep humans in the loop for plan approval and to get unblocked> Advancements in foundation models and simulation/distillation of small agent models based on reasoning traces + agent trajectories> CoV to minimize hallucinations for RAG> ADAS + AutoToS for dynamic agentic tooling that continuously evolves from accumulated knowledge> Runtime GoT + layers of AoT using MCTS for fetching relevant networks of thought> Build-time + offline training of MCTS policies via consolidation of layered GoT/AoT> Layered GraphRAG with domain-specific LGGMs for long-term memory with causal grounding> Optimizable agentic workflow graphs + simulation/self-play with tools inspired by DSPy, CLIN, MedAgent-Zero and agent symbolic learningIn unison, these design patterns may bridge adaptive reasoning + turing complete planning via search trees + dynamic agent definition/tooling that continuously evolves from accumulated knowledge with a balance of exploration/exploitation. Likely powered by at least 4 forms of continuous learning memory systems for:> Reasoning via “questioning” (Dynamic self-discover modules)> Reasoning via “thinking” (MCTS/GoT/AoT)> Planning (AutoToS/ADAS/LDB/LLM Compiler)> Long-term memory / user preferences (GraphRAG w/ CoV & tuned LGGMs)Links down below to relevant research 👇ADAS: https://lnkd.in/eVctVcETToS: https://lnkd.in/epDrsQSXGoT: https://lnkd.in/eUxg7_kzMindMap + GoT: https://lnkd.in/eHzxbHBpAoT: https://lnkd.in/ebP6irjJCoV: https://lnkd.in/eq7rMzm6Automated ToS: https://lnkd.in/emjDi-pYMCTS: https://lnkd.in/e4Yi_jjkCLIN: https://lnkd.in/eMq3TSC7Agents as Optimizable Graphs: https://lnkd.in/exD97gVrLGGMs: https://lnkd.in/eFkAHPNuGraphRAG Survey: https://lnkd.in/eWPYFxJAGoT/ToT Survey: https://lnkd.in/esP2fMiZ

4

AI agents will soon design themselves. Meta agents create "child" agents and evaluate their performance + refine their approach using an archive of previous discoveries ✨Let’s take a closer look into this emerging research...Summary & Key Takeaways 🧐> What are Meta Agents: These are agents that iteratively program new agents in code. The meta agent evaluates "child" agents based on their performance on tasks and refines their approach using an archive of previous discoveries (kind of like CLIN + MCTS)> Programming Language as Search Space: By defining agentic systems in code, the ADAS approach leverages the “turing completeness” of programming languages - enabling the discovery of any possible agentic system> Transferability & Robustness: These agents not only excel in their original domain but also perform well when transferred to different domains, tasks and modelsFuture Directions 🚀> Higher-Order ADAS: Self-referential ADAS where the meta agent itself can be improved through ADAS> Multi-Objective ADAS: Integrating multiple objectives (performance, cost, latency) into the ADAS search process> Complex Domains: Extending Meta Agent Search to more complex, real-world applications involving multi-step interactionsKey Insights ✨> ADAS shows parallels with CLIN, MCTS + IBM & Cornell University’s Automated Thought of Search, where code generation is used for programmatically defining search space - leading to more adaptable + general purpose agentic systems> ADAS framework + Automated ToS + CLIN + dynamic self-discover for semi-deterministic questioning of problem spaces + GoT/AoT/CoV with MCTS to find the best network of thoughts + MSFT's GraphRAG with LGGMs for layered graphs of memories, generalizations, thoughts and MCTS/ToS/ADAS policies may emulate the recursive and compound nature of human reasoning itself> These patterns combined with agentic workflow UX for self-play/simulation via DSPy and orchestration via optimizable graphs with assignable tools + deterministic workflows may bridge long-term memory, reasoning and turing complete planning with search trees and agents that continuously evolve based on accumulated knowledge with causal grounding - leading to emergent strategies that are difficult to achieve through purely hand-designed / statically defined control flow for agentsFor more on this, check out these papers 📑👇> Hu et al. (Aug 2024). "Automated Design of Agentic Systems"> Cao et al. (Aug 2024). "Automating Thought of Search: A Journey Towards Soundness and Completeness"> Zhang et al. (Jun 2024). “Accessing GPT-4 level Mathematical Olympiad Solutions via Monte Carlo Tree Self-refine with LLaMa-3 8B: A Technical Report”> Katz et al. (Apr 2024). "Thought of Search: Planning with Language Models Through The Lens of Efficiency"> Majumder et al. (Oct 2023). "CLIN: A Continually Learning Language Agent for Rapid Task Adaptation and Generalization"

5

Flow engineering is the future, not just prompt engineering... would be great to see more user-friendly UX around this concept!Reference this paper for more on this: > Sun et al. (June 2024). "Prompt Chaining or Stepwise Prompt? Refinement in Text Summarization"https://lnkd.in/e_EXjGPJ

Automating Thought of Search: A New Breakthrough in AI Planning by IBM & Cornell University ✨Planning is one of the most complex challenges with AI agents. Traditionally, this involves searching for a sequence of actions that lead from an initial state to a desired goal. Thought of Search (ToS) is a method proposed back in April 2024 by IBM which leverages LLMs to generate code that defines the search space - specifically focusing on creating a successor function and a goal test. While ToS required human experts to provide feedback to ensure its success, this new innovation removes humans all together 👀Key Takeaways> Researchers introduce AutoToS - an automated extension of ToS that removes the need for human intervention with automated feedback mechanisms used for guided generation of search components that are both sound and complete. This means that the functions produced by the LLMs accurately represent all necessary transitions and goal conditions without missing any steps or introducing errors.> AutoToS is a significant leap forward, achieving 100% accuracy across all tested domains, including BlocksWorld, PrOntoQA and Sokoban - all using unit tests and domain-specific checks to provide feedback to the LLM> The system requires minimal feedback iterations - comparable to what would be expected if a human were providing the feedback.Design patterns like AutoToS may play a key role in agentic workflows for planning and decision-making. While robust, these systems will likely still need to integrate more deterministic SOPs / workflows - perhaps akin to a hybrid approach, where deterministic workflows are supplemented by automated refinement processes with AutoToS to enhance flexibility and adaptability.For more on this check out the paper that was just released today on this and the original ToS paper:> Katz et al. (Apr 2024). "Thought of Search: Planning with Language Models Through The Lens of Efficiency"> Cao et al. (Aug 2024). "Automating Thought of Search: A Journey Towards Soundness and Completeness" 👇https://lnkd.in/emjDi-pY

2

Aidan Gomez (Cohere CEO) provides some key insights into how AI models are evolving and outpacing the needs of the average knowledge worker. Here are some key takeaways… > Each year, models become 10-100x cheaper to produce due to better data and cheaper compute> Each individuals ability to distinguish between generations of AI models becomes much harder because most of us are not experts in every field. The average person interacting with AI can’t fully absorb just how big of an intelligence leap there really is.> Therefore, models are starting to advance in much more specific capabilities - proving to be more useful not just for consumers but rather for scientific research & developmentCheck out the full interview here:https://lnkd.in/eMms64sS

2

How will we ever know who’s real when AI agents advance in the coming years?Imagine investing hours of time and emotion into what feels like a genuine connection - weeks of late night talks, shared interests, gaming sessions and intimate discussions - only to discover that the person you thought you knew was never real. Instead, you were interacting with a highly advanced AI agent - carefully designed to entertain and eventually manipulate you. This isn’t science fiction - it’s the future of the internet and it’s already starting to happen 🍓👀OpenAI, Microsoft, Harvard, University of Oxford, a16z, MIT and many other leaders in AI research have recently joined forces to combat this growing threat - proposing a solution called Personhood Credentials (PHCs). It's a lot like a digital drivers license issued to you by a central authority, allowing you to verify your humanity without exposing your name, location or any other identifying details to anyone on the web - all powered by zero-knowledge proofs.While PHCs could be the key to preserving trust online, they come with their drawbacks....> Privacy vs. SecurityIf not carefully implemented, PHCs could become effective tools of surveillance - allowing governments and corporations to track online activities at scale.> Impersonation & FraudEven with the best intentions, PHC systems could be vulnerable to subversion. Malicious actors may find ways to forge or steal credentials - undermining the system’s integrity and allowing them to impersonate humans even more effectively.> Physical ImpersonationAI agents can already talk like us, look like us and even execute sophisticated actions across various digital environments - making them indistinguishable from a real person. It is only a matter of time until these capabilities become cost efficient to run in real-time which will result in millions of digital scammers impersonating our friends and family members...Sooner rather than later, I really do think that we'll need to revive old techniques, like the call-and-response systems used during WW2 to distinguish friend from foe. These "challenge and password" or "challenge and countersign" methods, once critical in the chaos of the battlefield, may soon find new relevance in this new chapter of humanity vs AI.For more on this, check out the full paper: > Adler, Hitzig, Jain et al. (Aug 2024). "Personhood credentials: Artificial intelligence and the value of privacy-preserving tools to distinguish who is real online"

5