Talk by Adrian Bertagnoli (Callosum) arguing that the GPU-cluster era of identical hardware and monolithic models is giving way to heterogeneous intelligence, which optimizes model architectures, chip types, and workflows together. Covered results include a Qwen 3 VL 8B + Kimi K2.5 mixture that beat state of the art on Video Web Arena (18% and 25% above leading GPT/Gemini models at 3.7x lower cost and 3x faster) by routing zoom and visual parsing subtasks to a smaller model (11x speed and 43x cost improvement on those steps), and recursive long-context reasoning run on Cerebras, cutting cost 7x and latency 5x while matching accuracy. Callosum builds an automation layer that routes tasks to the right chip and model without per-subtask bespoke decisions. # Video Identity Extraction: Adrian Bertagnoli, Callosum ## Video Format & Context - **Format**: Conference talk with live Q&A and audience interaction ("who's here heard of recursive language models? Okay."). - **Speaker**: Adrian Bertagnoli, self-described **founding engineer at Callosum** ("I'm a founding engineer at Colossyan"). - **Topic framing**: Explicit talk structure stated upfront — why heterogeneity matters, how it's used today, and future scaling: > "why we care about heterogeneity in the first place, what particular aspect make it very conducive for scaling AI, um how it is actually used in practice today, and how we can use utilize it in the future." ## Explicit Self-Description & Creator Claims - **Company mission claim** (creator claim about new hardware unification): > "currently, there's no interface which allows it... this new hardware to be unified and constructively help the current compute stack. And so, this is what we aim to change." - **Formalization claim**: "the benefit is not simply a belief that we have. We actually formalize it and proved it mathematically" — the **principle of maximum heterogeneity**, claimed to hold across "neuroscience, economics, and ecology." - **Institutional validation claim**: "We are actually working with Aria, the UK Institute. Um we got a 3 million grant for operating the first heterogeneous collocated cluster in the UK." - **Recruitment signal**: "we are hiring" — closing line self-description. ## Demonstrated Scope (Observable Patterns) - **Technical depth**: Discussion of mixture-of-experts, multi-agent systems, prefill/decode disaggregation, recursive language models (citing "a seminal paper that came out of MIT last October"), benchmarks (ULong, Video Web Arena). - **Product/engineering demonstration**: Presents concrete benchmark results with named hardware and model partners: - Cerebras: "seven times cheaper and five times faster" vs GPT-5.2 - SambaNova: "12 times cheaper and three times faster" - Video web navigation: beating "GPT 5.2 and Gemini 2.5 by 18 and 25% respectively" using mixtures (Qwen 3 VL-8B-Instruct, K2.5) - **Terminology creation**: Coins/uses internal vocabulary — "heterogeneous recursion," "multi-agent heterogeneous intelligence," "production function," "principle of maximum heterogeneity." - **Evolution of system disclosed in Q&A**: "initially we started doing bespoke decisions... but since then, we have created an automation layer that detects the task complexity and automatically predicts the best model." ## Intended Audience - **AI researchers and infrastructure engineers** familiar with frontier models, scaling laws, and hardware stacks; assumed baseline knowledge (e.g., unexplained references to context rot, needle-in-haystack complexity classes, MoE). - **Potentially investors/hires** — given the funding claim, hiring pitch, and competitive positioning against Nvidia's paradigm: > "The second era was making compute massively parallel. This is dominated by Nvidia. And the third paradigm, compute is going to become heterogeneous." ## Claims vs. Demonstrated Evidence Distinction - **Creator claims** (not independently verified in-video): mathematical proof of heterogeneity benefit; ARIA £3M grant; benchmark figures. - **Observable demonstrations**: structured technical argumentation, working knowledge of current model/hardware landscape, live Q&A handling indicating domain fluency. ## Channel-Level Presentation Signals (Video-Level Only) - Positions the channel/organization at the **AI infrastructure orchestration frontier**, contrasting itself with the "homogeneous" Nvidia-era scaling paradigm. - Tonal pattern: promotional-but-technical — pairs quantitative claims ("11 times faster and 43 times cheaper") with explicit thesis framing ("This is the worst our infrastructure will ever be.") # Delivery Observations — Adrian Bertagnoli, "Scaling the Next Paradigm of Heterogeneous Intelligence" ## Tone & Overall Style - **Explainer-style conference talk**: measured, pedagogical tone with a clear roadmap up front: *"I'm going to start um with explaining why we care about heterogeneity in the first place... how it is actually used in practice today, and how we can use utilize it in the future."* - **Conversational fillers throughout**: frequent "um" and "uh" markers appear across the talk, e.g. *"this becomes less and less relevant. So, it's already changing um"* — suggests unscripted or lightly rehearsed delivery. - **Enthusiasm surfaces at results moments**: *"so it's basically like having your cake and eating it, too"*; *"This is the worst our infrastructure will ever be."* ## Recurring Linguistic Patterns ### "So" as a discourse anchor - Used heavily to open new points: *"So, homogeneous intelligence in in terms of AI mainly refers to..."*; *"So, why heterogeneity? Why is it a good thing in the first place?"*; *"So, here are results."* - This is the most consistent verbal tic in the transcript, appearing at nearly every section transition. ### Self-corrections and restarts - *"heterogeneia heterogeneity"* — stumbles on his own central term. - *"we we acknowledge that the problem is heterogeneous uh it itself"* - *"Kimi Kimi alone"*; *"uh 2.5 and GPT 5.2"* - Repetition for emphasis: *"It it decomposes into"*, *"um uh acting together"*, *"and and that's why"*. ### Named coinages / branded terminology - Repeatedly introduces terms with framing phrases: - *"something we like to call multi-agent heterogeneous intelligence"* - *"what we like to call a production function"* - *"a primitive we like to call heterogeneous recursion"* - The pattern **"something/what we like to call X"** appears at least three times — a stable teaching habit. ### Triadic structures - Repeatedly organizes ideas in threes: - Architecture / workflow / hardware levels: *"On the architecture level... On the workflow layer... And finally, on the hardware level..."* - Closing framing: *"models, workflows, and silicon co-evolve"* and *"smarter, faster, and cheaper."* ## Teaching Habits - **Audience calibration checks**: asks the room directly — *"who's here heard of recursive language models? Okay. So, for those of you who um know don't know..."* — then adapts by explaining the MIT paper before building on it. - **Intuition-first explanations**: *"to give you an intuition about what I mean with heterogeneous intelligence, I want to take a step back"* — before formal claims. - **Concrete analogy for complexity classes**: explains O(1) vs O(N) with a plain-language example — *"if you're doing a needle in a haystack task, that is O of 1"* and *"Adding up the rows would be O of N."* - **Numbers-heavy evidence delivery**: results stated as paired speed/cost multipliers: *"seven times cheaper and five times faster"*, *"12 times cheaper and three times faster"*, *"18 times cheaper than using uh GPT 5.2 alone"*, *"11 times faster and 43 times cheaper."* - **Deictic gestures toward visuals**: relies on slides — *"on the right on the left, you see a figure"*, *"So, here are results. Um you can see this is on the Ulong benchmark."* (Visuals not verifiable from transcript; only the verbal references are observable.) ## Communication Patterns - **Problem → principle → practice arc**: motivates (*"Why heterogeneity?"*), formalizes (*"we formalize it and proved it mathematically"*), then demonstrates (*"So, how do we use this in practice?"*). - **Anticipates skepticism**: *"the benefit is not simply a belief that we have. We actually formalize it"*; and reiterates: *"I've been telling you about the benefits of heterogeneity, but I've not told you anything about what it actually means in terms of AI."* - **Direct Q&A handling**: answers the one audience question concisely and honestly about evolution of the method: *"initially we started doing bespoke decisions... but since then, we have um created an automation layer."* - **Practical close**: ends with identity, invitation, and a hiring pitch: *"My name is Adrian Berczynski and if anyone is interested, uh we are hiring."* ## Uncertainty Notes - **Name inconsistency**: the speaker opens with *"My name is Adrian Bertagnoli"* and closes with *"My name is Adrian Berczynski"* — at least one is a misspoken variant; the title metadata says "Bertagnoli." Cannot determine which is correct from delivery alone. - **Filler frequency and pacing** are inferred from transcript markers only; actual delivery cadence, gestures, and slide content are not observable here. - The "we like to call X" pattern is well-supported (3+ instances); the triadic structure pattern rests on ~3 examples and may be partly incidental. # Observation Record: Scaling the Next Paradigm of Heterogeneous Intelligence **Speaker:** Adrian Bertagnoli (also stated his name as "Adrian Berczynski" at the end of the talk), founding engineer at Callosum (transcribed as "Colossyan"). ## Core Thesis - The current paradigm of **homogeneous intelligence** — scaling single models on fleets of identical chips — arose from neural scaling laws (more data and parameters → better models), but this is rooted in the **training domain** and becomes less relevant as the field moves toward **inference**. - Heterogeneity is beneficial, and Bertagnoli claims this is formalized and "proved mathematically," holding across **neuroscience, economics, and ecology** — "under any reasonable amount of constraints, heterogeneous systems outperform homogeneous ones." ## Stages of Heterogeneity (as presented) 1. **Mild heterogeneity** (current state): homogeneous clusters with prompt variety; multi-agent systems using different LLMs per sub-agent; mixture-of-experts. 2. **Greater heterogeneity**: different chips for different models (different GPUs), and diverse model types (state-space models, diffusion models) interacting on optimal hardware. 3. **Full heterogeneous paradigm**: co-evolution of hardware and software — "complete vertical integration of intelligence and hardware." ## Why Heterogeneity (stated reasoning) - Real-world problems are **complex, multi-step, open-ended**, and decompose into sub-problems requiring "vastly different types of intelligences"; scaling a singular type is "very inefficient." - New silicon generations are arriving but lack an interface to unify with the compute stack — this is what Callosum "aims to change." - **Principle of maximum heterogeneity**: agents' skills form distributions over a skill space; a given "production function" suits one problem's demand but not another's. Homogeneous scaling either matches one peak or yields "generalists" forming a "very short cylinder" that doesn't meet the production function. ## Practice: Optimization at Three Layers - **Hardware**: choose hardware per agent's computational demands. - **Agent layer** and **workflow layer**: described via two examples below. ### Heterogeneous Recursion - Based on the recursive language models paper "out of MIT last October": context rot depends on information complexity — O(1) tasks (needle in a haystack) scale fine; linear/quadratic tasks degrade at "around 60 to 30%" context occupancy. - Recursive LMs treat context as an **environment**: a coding agent interacts with a context file via Python REPL (keyword search, regex), extracting sub-contexts passed to identical recursive agents. - Callosum's extension: map sub-contexts to **different chips and models** to emulate performance more cheaply. - **Results (Ulong benchmark)**: GPT-5.2 (~2000s, ~$3.75/task) vs. their system on **Cerebras: 7× cheaper, 5× faster**; on **SambaNova: 12× cheaper, 3× faster** (trading some latency for price). ### Visual Web Navigation (Video-Action Language Models) - Used a mixture of open and closed video-action language models; beat Video Web Arena state-of-the-art, **outperforming GPT 5.2 and Gemini 2.5 by 18% and 25% respectively**. - Reasoning: the problem itself is heterogeneous — decomposing into visual and textual reasoning steps requiring different models. - Pareto-frontier claims: - Qwen 3 VL 8B-Instruct (transcribed as "Quant 3 VL8B-Instruct") + Kimi K2.5: **1.3× faster than Kimi alone, 18× cheaper than GPT 5.2 alone**. - Qwen 3 + GPT: **3× faster, 3.7× cheaper**. "There's no downside." - Subtask offloading (e.g., zooming to less intelligent models — "you don't need GPT to zoom"): **11× faster, 43× cheaper than ChatGPT** on those subtasks. ## Future of Compute (three eras, per Bertagnoli) 1. **CPU era**: compute got quicker. 2. **Nvidia-dominated era**: massively parallel compute. 3. **Emerging era**: heterogeneous compute mapped onto multi-agent workloads across different chips. - Announced a **£/$3 million grant from ARIA, the UK Institute**, for the "first heterogeneous collocated cluster in the UK." ## Closing Claim > "This is the worst our infrastructure will ever be." — every new source of diversity makes the system "smarter, faster, and cheaper." ## Q&A - **Question**: Is subtask-to-model mapping (e.g., zooming) hardcoded? - **Answer**: Initially bespoke mapping decisions; since then an **automation layer** detects task complexity and predicts the best model and hardware. ## Notes and Unresolved Tensions - Speaker's name appears as both "Adrian Bertagnoli" (intro) and "Adrian Berczynski" (closing) in the transcript — a transcript inconsistency preserved as stated. - Company name appears as both "Colossyan" and "Callosum" (per title metadata) — inconsistency preserved. - The mathematical "proof" of heterogeneity's benefit is asserted, not shown in detail; the cross-domain validation is claimed, not evidenced in the talk. - Benchmark figures (GPT-5.2, Gemini 2.5, Kimi K2.5, Qwen 3) are speaker-reported claims without cited methodology beyond the benchmarks named. # Attribution Record: "Scaling the Next Paradigm of Heterogeneous Intelligence" ## Speakers & Roles | Speaker | Role | Notes | |---|---|---| | **Adrian Bertagnoli** (also self-identifies at the end as "Adrian Berczynski") | **Creator / Presenter** | Founding engineer at "Colossyan" (spoken company name; title metadata says "Callosum"). Delivers the full talk. | | **Unidentified audience member** | **Audience / Questioner** | Asks one Q&A question about task-to-model mapping. | > Note: The transcript contains inconsistent self-identifications ("Adrian Bertagnoli" at start, "Adrian Berczynski" at end) and inconsistent company names ("Colossyan" spoken vs. "Callosum" in the title). Both spellings are preserved as spoken; uncertainty remains. ## Provenance & Referenced Sources - **Recursive language models paper**: Cited as a "seminal paper that came out of MIT last October"; no author names or title given. - **Benchmarks referenced (no documents cited)**: - **Ulong benchmark** (from the recursive LM paper, per speaker) - **Video web arena** benchmark - **Third-party models/products mentioned**: GPT-5 / GPT 5.2, Gemini 2.5, Kimi K2.5 (spoken as "Kimik A 2.5"), Quant 3 VL8B-Instruct, ChatGPT, Cerebras, SambaNova, Nvidia (era framing). - **Funding claim (speaker's own claim)**: Working with **ARIA, the UK Institute**; a **£/$3 million grant** for "the first heterogeneous collocated cluster in the UK" — stated as "3 million" without currency. ## Performance Claims (speaker-asserted, unverified) - Cerebras: 7× cheaper, 5× faster than GPT-5.2 on Ulong; SambaNova: 12× cheaper, 3× faster. - Beat state-of-the-art on video web arena: +18% vs GPT 5.2, +25% vs Gemini 2.5. - Mixtures: 1.3× faster than Kimi alone; 18× cheaper than GPT 5.2 alone; Quant 3 + GPT: 3× faster, 3.7× cheaper; subtasks 11× faster, 43× cheaper than ChatGPT. - Mathematical formalization of heterogeneity benefits across neuroscience, economics, ecology — described but no citation provided. ## Promotional Segments (Flagged) - **Hiring pitch** at the close: "we are hiring" — explicit promotion. - **Company/product promotion** throughout: results framed as "our system," positioning the company as spearheading the heterogeneous paradigm; **grant announcement** serves promotional framing. - Q&A answer describes the company's automation layer as a differentiator. ## Availability & Metadata Limits - **No curated metadata supplied** — no verified speaker bios, citations, comment availability, or publication details. - All quotes, figures, and citations come solely from the unverified transcript. - Comment availability: **unknown/not stated**. Thank you for coming to my talk. My name is Adrian Bertagnoli. I'm a founding engineer at Colossyan and today I'm going to be talking about scaling the next paradigm of heterogeneous intelligence. So, I'm going to start um with explaining why we care about heterogeneity in the first place, what particular aspect make it very conducive for scaling AI, um how it is actually used in practice today, and how we can use utilize it in the future um to actually scale the next paradigm of intelligence. So, to give you an intuition about what I mean with heterogeneous intelligence, I want to take a step back and um explain the current prevailing paradigm of homogeneous uh intelligence. So, homogeneous intelligence in in terms of AI mainly refers to scaling single models on a fleet of identical chips. So, this was largely this era was largely brought about by the discovery of neural scaling laws, which showed us that more data and more parameters leads to better models. However, this is primarily rooted in a training domain and while we move towards an inference domain, this becomes less and less relevant. So, it's already changing um and we already see some level of heterogeneity in um going into our current systems. So, on the architecture level, we see that mixture of experts are replacing large dense models. On the workflow layer, we see that single LLM calls are being replaced by uh multi-agent systems. And finally, on the hardware level, single chips are being replaced by pre-fill decode disaggregated systems. So, given that we are currently at the state of mild heterogeneity, how can you imagine um a greater level of heterogeneity? How will that appear? So, initially, we'll be what we currently are experiencing is mild heterogeneity. So, everything is still running primarily on on homogeneous clusters, but we have some variety in the prompts. When we run uh multi-agent systems, we might use different LLMs for different sub-agents. Um again, we have mixture of experts. Uh when we increase the heterogeneity, we might start to use different chips um for different models. So, different LLMs might be put on different GPUs. Uh they might be interacting. We might be using different models completely. So, we have a increase of state-space models, diffusion models, all interacting with each other, all on optimal hardware that exists currently. In the last stage, uh where we really see the the heterogeneous paradigm unfolding, is when we have a co-evolution of systems uh hardware and software. So, there will be a unification where you'll have um a complete vertical integration of intelligence and hardware. So, why heterogeneity? Why is it a good thing in the first place? So, real-world problems are complex, multi-step, and open-ended. They decompose into sub-problems, which require vastly different types of intelligences. So, scaling a singular type of intelligence to solve these is very inefficient and and not optimal. So, how do we solve them? Solving these actually requires models of different architectures and sizes working together um uh acting together in long horizons, something we like to call multi-agent heterogeneous intelligence. Furthermore, new generations of silicon is coming towards the market, but currently, there's no interface which allows it to um this new hardware to be unified and and constructively um help the current compute stack. And so, this is what we aim to change. So, heterogeneia heterogeneity, the benefit is not simply a belief that we have. We actually formalize it and proved it mathematically um on the right on the left, you see a figure outlining uh the principle of maximum heterogeneity. So, these are heterogeneous agents, where the color indicates um a distribution over a skill space. Um if you take if you have a communication between these, here indicated by a ring topology, you can have a what we like to call a production function. And the production function is simply the demand uh can can be well suited for the demand of one problem, but ill suited for another problem. So, here we have a a production function that's well suited for demand A and ill suited for demand B. If you want to do this in a homogeneous fashion, you would either be able to only scale one peak or in the optimal case to match this demand function, you'd have only generalists, so as broad as possible the skill set, but then ultimately you'd have a very short cylinder that does not meet the production function readily. So, we formalize this and we um saw that across many domains, including neuroscience, economics, and ecology, these these trends hold and under any reasonable amount of constraints, heterogeneous systems outperform homogeneous ones. So, how do we use this in practice? Like, I've been telling you about the benefits of heterogeneity, but I've not told you anything about what it actually means in terms of AI. So, we optimize um multi-agent systems at three different parts of the workflow. So, all the way from the hardware, where agents run on, we choose different hardware depending on the computational demands on the agents, and then how agents interact and what workflow they construct. Um so, we have already demonstrated multiple uh benefits of of uh this type of orchestration and I want to go into a couple ones um main namely in the workflow, something a primitive we like to call heterogeneous recursion, and in the agent layer, I want to talk about multimodal multimodal video action language models. So, heterogeneous recursion. This is something um who's here heard of recursive language models? Okay. So, for those of you who um know don't know recursive language model, it's kind of a seminal paper that came out of MIT uh last October. And they basically showed that even if you only occupy a small um percentage of the context window, you still can have dramatic context rot depending on the information complexity you want uh from the prompt. So, if you're doing a needle in a haystack task, that is O of 1. The the information requirement scales is constant throughout regardless of how big the prompt is. And then you can imagine adding up the rows. You you you give uh rows and columns. Adding up the rows would be O of N, because as the prompt increases, the inform- informational requirement increases linearly. So, if the if you have a constant information requirement, it scales well. You you can occupy the full context window and actually get a a good answer. However, when you go to linear or quadratic, uh it degrades at around 60 to 30%. So, recursive language models solve this problem by actually treating the context as a um environment, rather than putting it all into the prompt. So, in practice, this looks like you present the context in a file, and then the a coding agent interacts with it programmatically through Python REPL, um basically doing keyword searches, regex, and other tricks to extract sub-context, and this sub-context is then passed off to an identical recursive agent. So, this agent then can answer the question or spawn another recursive agent. And and that's why it's called recursive uh language model. So, we simply extended this concept um instead of using a single model on a single chip, we map based on the sub-context generated towards different chips and different models to emulate the performance while drastically um being cheaper and faster. So, here are results. Um you can see this is on the Ulong benchmark. This is basically um the benchmark they used in the paper um and GPT-5 GPT-5.2 was the um most recent one when we produced this work uh sits around here, where it takes around 2,000 seconds um to run through the the benchmark and it costs around uh $3.75 for one task. Our system, when we go on Cerebras, we are seven times cheaper and five times faster. So, you save incredibly much time, are a lot cheaper, so it's basically like having your cake and eating it, too. Um with SambaNova, we even get uh further. We push the price down even further at the cost of some latency. So, we're 12 times cheaper and three times faster. So, these are like making architectural decisions that are not like simply based on the hardware. You can make huge impactful um price differences and and and while emulating the intelligence you would have from frontier models. So, the next problem we wanted to address is basically um visual web navigation. So, we used a mixture of open and closed uh video action language models. Um and we managed to beat uh the state-of-the-art of video web arena beating GPT 5.2 and Gemini uh 2.5 by 18 and 25% respectively. And not only this, the way we did it is instead of treating the problem as a homogeneous one, we we acknowledge that the problem is heterogeneous uh it itself. It it decomposes into multiple steps of visual reasoning of of textual reasoning and each of these subcomponents requires different models to be um completed successfully. So, here you see a fundamental shift of the Pareto frontier where you see singular models like Kimik A uh 2.5 and GPT 5.2 are outperformed by mixture, a heterogeneous set of of models. So, when we use Quant 3 VL8B-Instruct and Kimik A 2.5, we're 1.3 times faster than using Kimi Kimi alone. We're 18 times cheaper than using uh GPT 5.2 alone. Um and if we use uh Quant 3 um plus GPT, we're actually three times faster and 3.7 times cheaper. So, this is only benefit. There's no downside uh in in constructing this in a heterogeneous manner. So, one part of of our differentiating factor, why how we were able to beat the state-of-the-art, is that we mapped certain subtasks like zooming and and creating a different visual um reasoning for the agent, we offloaded that into less intelligent models because you don't need GPT to zoom for you. So, alone on these subtasks, we're able to be 11 times faster and 43 times cheaper than using ChatGPT. And so, this is what overall overall accumulates towards these 3.7 times cheaper and three times faster. So, looking ahead, how do we view the future of compute? The first era of scaling compute was dominated by the CPU where compute got quicker. The second era was making compute massively parallel. This is dominated by Nvidia. And the third paradigm, compute is going to become heterogeneous mapping onto multi-agentic workloads and optimally um mapping these workloads onto different chips. We are actually um working with Aria, the UK Institute. Um we got a 3 million grant for the first for operating the first heterogeneous collocated cluster in the UK. So, we really want to make a difference and and spearhead this new era of innovation. So, the era of homogeneous scaled delivered extraordinary progress. We should be grateful for it. What comes next is heterogeneous intelligence where models, workflows, and silicon co-evolve and every new source of diversity makes the whole system smarter, faster, and cheaper. This is the worst our infrastructure will ever be. Thank you. How do you define which task to run on the faster uh cheaper model? Like for instance, is zooming that something that's hardcoded like oh, if you have to zoom, use this model or if you have like the smarter model, use that type of model? So, it is um initially we started doing bespoke decisions on mapping uh certain simple subtasks to simple models, but since then, we have um created an automation layer that detects the task complexity and automatically predicts the best model uh the best suited model and hardware. Any other questions? Great. Thank you so much for your attention. My name is Adrian Berczynski and if anyone is interested, uh we are hiring, so yeah. Great. Thank you very much.