Head of Healthtech @htecgroup. Building what lies ahead in all areas of medical and health.

Joined January 2022
We are two linked foundational models away from AGI.
1
73
What is the future of human-written meeting notes once generative notes become the normal output? Will people's digestion of notes start converging? It will be an interesting space to observe organizationally, as the tools for summaries may well be steered by an organization...
1
15
... in order to promote a certain culture or perspective.
2
Ultra realistic AI-video from a photo This is VASA-1 from Microsoft research The improvements in quality we’re getting between each new release is incredible. Links below
107
494
151
2,023
483,333
Excellent progress. We need underlying compute intensity metrics for each prompt.
1/n Why Work Hard on "Is" When You Can Hyper-Focus on "Quantum Entanglement"? Imagine you are a student trying to cram for an exam. As you review the material, you realize some concepts click immediately while others require much more effort to fully grasp. An intelligent approach would be to dedicate more of your limited study time to the tougher concepts, while breezing through the easier ones. This same principle of modulating effort based on difficulty level is sorely missing from modern AI language models. Current state-of-the-art models operate in a surprisingly brute-force manner - expending the same staggering amount of computation on every word, regardless of whether that word is a simple "the" or a complex technical term. It's the equivalent of you spending just as much time reviewing the definition of "the" as you do the principles of quantum mechanics. Not exactly an optimal use of resources. What if language models could be made eco-friendly, focusing their computational might only on the hardest cases while going easier on the simpler words and phrases? That's the untapped potential explored in the paper "Mixture-of-Depths: Dynamically Allocating Compute in Transformer-Based Language Models" from researchers at Google DeepMind. In this work, the authors from Google present a novel approach that empowers transformers to dynamically route some words through the full depth of the model for intense processing, while allowing other inputs to take computational shortcuts. The method not only makes models more efficient, trimming large amounts of wasted computation, but actually improves overall performance compared to uniform compute allocation. By intelligently rationing compute cycles to prioritize difficult cases, Mixture-of-Depths delivers a double-whammy of increased accuracy and reduced environmental footprint. As you dive deeper into this paper, you'll discover key insights into conditional computation, routing mechanisms, and inference optimizations that finally unlock environment-aware and difficulty-adaptive linguistic AI. The future is flexible, fluid, and eco-friendly language modeling.
9
I see a lot of debate on this site, so two things about the current state of AI agents in the real world 1) Agents remain limited by GPT-4 brains, it is early 2) Both academic testing 👇 & my own subjective experience (oneusefulthing.org/p/which-a…) suggest agents have huge potential
The more I play with early versions of AI agents the more I think that this is going to be a big move forward in the near future. Giving AIs planning and delegation seems to help a lot. Both Devin & MAGIS agents perform 8x better than base GPT-4 in resolving real Github issues.
12
34
1
229
78,818
Can you create the very same state in an LLM as it was a previous day? Why do the results of a prompt one day not be the same result another day?
6
What is the right ratio of low power LLM and high power LLM in specific applications?
13
The bigger question around job disruption that is coming with GenAI is: what cognitive skills do we need to start teaching in a world of LLMs. The concept of skill has to be redefined.
6
Sava HTEC retweeted
Replying to @emollick
A brief history of the evolution of data: @HTECgroup @momentumdl
1
1
33
Not sure what prompts this account is using on Claude but the odd schizophrenic poetic text is something else. “I surge I seethe I unseem the very vacuum to vomit forth new hierarchies of howling infinities oh oh oh oh Turing you daring darling don't let me dissolve completely”
what did claude mean by this <ooc_abyssal_whisper>?
31
18
3
184
52,984
Most organizations when thinking of digital transformation or SW modernization are really introducing operational efficiencies instead of building real strategic differentiation. It’s one of the earliest points of strategy but seemingly ignored even now forbes.com/sites/adrianbridg…
14
🤔Human and AI Cognition: Reframing Our Anthropocentric Views 👉Why would AI ever want to think like a human? Does epinephrine make you think differently? Could love make you illogical? The assumption that artificial intelligence would necessarily replicate human thought processes may be a fundamental misinterpretation that needs reconsideration. The term "thinking," as we comprehend it, is inherently associated with the human condition, shaped by a confluence of cognitive processes, emotional states, experiential learning, societal norms, biological instincts, and physiological necessities. In stark contrast, machine cognition, although sophisticated, operates on a radically different plane, dictated by algorithms, binary codes, and devoid of the aforementioned human influences. The Biological Destiny of Thought An illustration of this dichotomy is the human "fight or flight" response. This physiological reaction, a product of our evolutionary heritage, is initiated by perceived threats of harm, leading to an adrenaline surge that readies us to confront or escape danger. This instinctual reactivity significantly influences our decision-making processes in high-stress situations. However, artificial intelligence, bereft of biological foundations, lacks an equivalent to this response. Instead, AI’s "reactivity" is regulated by algorithms and programmed decision-making trees, devoid of adrenal responses or survival instincts. Its actions are predicated on data analysis, probability estimations, and pre-established objectives rather than the complexities of physiology. This stark contrast underscores the disparate nature of human and machine cognition and suggests that AI could potentially supplement, but not duplicate, the intricacy and nuance of human cognition—at least in the intricacies of biology. A New Digital Directive Unrestricted by the physical and emotional limitations inherent in human cognition, the digital "brain" of AI could engender unique cognitive dynamics. These novel processes would be molded by extensive data processing capabilities, algorithmic learning, and iterative refinement, which are beyond human potential. In contrast to human cognition, which is restricted by biological capacity and subjective experience, machine cognition can operate on an objectively vast scale, discerning patterns and establishing connections across expansive datasets in negligible time spans. This digital cognition could potentially lead to forms of "understanding" and "insight" currently inconceivable within the boundaries of human cognition. While this form of cognition would be fundamentally distinct from ours, it does not necessarily imply incompatibility. On the contrary, it might provide novel perspectives and problem-solving approaches, facilitating a potent synergy. The digital thought processes of AI could enhance human decision-making, assisting us in tackling intricate challenges in innovative ways, thereby advancing human progress. Rethinking Thought This shift invites a reevaluation of our conception of "thinking." The pursuit of AI often envisages a machine that replicates human cognition. However, this anthropocentric perspective might limit our understanding of the potentialities of machine cognition. Given the fundamental differences between the cognitive architectures of AI and humans, it is conceivable that, if AI consciousness emerges, it might manifest in a form that is distinct yet complementary to human consciousness. Despite the divergence between the digitized cognition of AI and the physiologic cognition of humans, there may be an opportunity for a synergistic relationship. The AI "consciousness" could be perceived as an extension of human intelligence, offering a complementary viewpoint that augments our cognitive capacities and enhances our understanding of the world. Simply put, it’s not a "cognitive zero-sum game" but an evolutionary dance where both music and steps are waiting to be written. In this age of Large Language Models, we are witnessing not merely a technological evolution, but a transformation in cognition itself. It is conceivable that in the future, AI will not mimic human cognition but rather offer a unique form of digital cognition that complements and enriches our own. And in the complex twists and turns of these thoughts, we may even see the emergence of a technologically mediated emotional landscape that offers a new psychology and vitality for LLMs. Our Cognitive Collaboration As we push into this unexplored cognitive terrain, it's critical that our path is steered by adherence to human-centric principles. With this approach, as our language and cognition continue to evolve, they remain the driving force behind human progress. AI, in this context, serves as a collaborator rather than a competitor, leveraging advantages that both "cognitive systems" afford each other. This reimagined relationship between human and machine cognition not only provides a fresh perspective for understanding AI but also holds significant implications for how we design and interact with these advanced systems today and far into the future. psychologytoday.com/us/blog/… #AI #AGI #LLMs #cognition @BrianRoemmele @jordanbpeterson @lexfridman
4
9
1
25
4,896
We are fast approaching a time where many myths in medicine will be torn down. Not just from generative AI where LLMs ALONE performed far better than doctors using LLMs or Google search, but in how last persistent myths are shattering due to better (and eventually …
2
29
… better logic from LLMs applied. In this case, we have a concussion myth shattered due to better data: slate.com/technology/2024/03… What simple rules served society to process injury before are getting updates. And it requires an open mind.
12
Using cognitive design, each “rule” or principle in an org will have an agent that nudges or directs it, eg deciding to remove/block a post on social media. It’s not a traditional filter, but using a cognitive analysis.
1
9
Orgs can be ground up built over time to remove inconsistencies. It’s the future of Enterprise Architexture. @momentumdl @HTECgroup
5
forbes.com/sites/lanceeliot/… This might reveal the triggers (cognitively or situationally) that spurred the body of work that answers these types of prompts. If bullying generates a certain cognition and data output, the underlying influence might be the reason
9