Elissa Murphy

Thank Dani Cherkassky Great piece on the architecture shift voice AI needs. The core thesis: cloud-centric voice systems fail in real-world conditions because they are too slow, too expensive to keep always-on, and deaf to spatial context. The solution is a hybrid architecture—fast, always-on edge processing (Spatial Hearing AI and Cognition AI) handling 80% of interactions locally, with cloud LLMs reserved for complex reasoning. But the Kahneman thesis runs deeper than the article suggests. System 1 and System 2 are not just useful metaphors—they reflect an actual biological sequence. Fast, intuitive processing (System 1) evolved first; slow, deliberate reasoning (System 2) was layered on top. The brain's architecture reflects this: spatial-acoustic processing happens in milliseconds at the brainstem and primary auditory cortex, long before language centers engage. Kardome is not just borrowing Kahneman as marketing. Their architecture is neurologically correct—it replicates the hierarchical, parallel, spatially-aware processing that makes human auditory cognition so robust. Here is the deeper point: hearing preceded language in human evolution, and spatial hearing preceded both. The mammalian auditory system evolved to detect predators and locate prey—survival functions that required sub-second latency and continuous environmental awareness. Language processing was layered on top of this infrastructure. Current voice AI inverts this evolutionary logic: it starts with language models and treats spatial audio as an afterthought. Kardome restores the natural hierarchy. 🙌 Danny Shapiro, Rory Sutherland

36

2 Comments

Deep architecture innovation drives durable advantages in #AI compute for Mythic Also worth noting Mythic’s chips can be made on multiple friendly fabs in secure territories like the US, with at least three more levels of process geometries worth of headroom to exploit for their already massive SWaPC lead…

13

1 Comment

I sat down with Stuart Turley and Kyle Koss on Energy News Beat to talk about AI at the energy edge, bringing compute to where power already lives, so mid-tier operators aren’t locked out by mega-campus timelines and capital. The infrastructure piece is real. The part I care about most is what happens after the model speaks. When inference runs at the wellsite, the plant, or a modular edge node, speed is the easy win. The hard requirement is the same one we already insist on in financial services: can you defend the conclusion when someone paid to disagree asks? Explainability is non-negotiable. That is why we issue Decision Records, conclusion, supporting evidence, conflicts disclosed, confidence with basis, and named human judgment. Not a fluent answer. A reconstructable basis. We built reView as the trust layer of the AI stack, above the data and IT you already run. Financial services and energy first. Defense and intel are one lane, not the whole map. ENB: https://lnkd.in/gEk9JaZ8 Edge AI Governance: https://lnkd.in/g62ABs6W https://www.data2.ai #ExplainableAI #DecisionIntelligence #FinancialServices #Energy #TrustworthyAI #EdgeAI

22

1 Comment

Enterprise AI has a equipment problem, not a compute problem. I contributed to a new 3-part series on the Intel community blog: "From Gold Rush to Factory: How to Think About TCO for Enterprise AI." The core idea: most organizations are still in gold-rush mode — buying GPUs before understanding the work. But the enterprises getting real ROI from AI are thinking like factory operators. Match the equipment to the job. Build repeatable workflows. Optimize the CPU-to-GPU ratio by workload, not by hype. One story from the piece that still gets me: a genomics company deployed high-end GPUs to meet a seven-day SLA. Got results in under a minute. Impressive — until someone asked if a two-hour completion on CPUs would suffice. It did. The GPUs were decommissioned. That's not an edge case. It's closer to the norm than most want to admit. The constraint isn't compute. It's the lack of a system for matching the job to the right equipment. Part 1 is live now. Parts 2 and 3 are coming. https://lnkd.in/gQX_pe_2 cc: Rahul Awasthy Lynn Comp

6

1 Comment

We believe Nvidia retains its time to market advantage - and this, plus its "CUDA moat" should enable it to maintain leadership in the accelerator market for now given its rapid pace of innovation in the near term. Longer term, we see a growing presence of custom ASICs for internal workloads, and expect hyperscalers to selectively adapt their internal offerings for use by external customers. As raw compute performance reaches physical limits (and accelerators are already reticle limited), we expect further performance and cost improvements to be driven by innovation across networking, memory, and packaging. We see Nvidia and Broadcom as best positioned to leverage developments in these areas. Nvidia is significantly outspending competitors on R&D, has a strong positioning in networking with its Mellanox business, and has started to also deliver innovations in memory technology with its context memory storage controller offering. Broadcom also continues to lead the market with its best-in-class processor/accelerator solutions plus its industry-leading ethernet networking and SERDES capabilities.

1

Just back from TDK Ventures 100X 2026 Conference where I served on a panel. A few insights from the morning session. THE KEYNOTE: “100X is not an invention. It enables an ecosystem” -- Chuck Mattera Chuck Mattera (former CEO of Coherent, now CEO of Avalanche Thinking) argued that no 100X has ever been a single breakthrough, but a chain: transistor, IC, microprocessor, CMOS scaling, SoC, AI,. The question isn't what you invented — it's what becomes possible to compound. Building an ecosystem is like playing Go: the board is visible to everyone, the opportunities are not. And manufacturing is where compounding starts: making things creates learning, scale compounds learning, learning compounds speed. Bring manufacturing back and you build the ecosystem that learns fastest.  I couldn’t agree more. MY PANEL: FROM MODELS TO MOTION — AI ENTERS THE PHYSICAL WORLD Moderator: Modar Alaoui, Panelists: Les Karpas (NVIDIA Inception), Claire Delaunay (OPALIN), Ben Burchfiel (Walden Robotics) I think my panel agreed on more things than we disagreed on. There seemed to be consensus that there is no single winning embodiment.  The moat is shifting to whoever owns a data engine (which is not the same as owning data) and whoever can learn from the customer feedback cycle. I shared where I'd invest for the next decade: data engines and the application-specific robotic foundation models under them. In hardware, components like actuators, hands and tactile sensing. Models move fast; their advantage decays. Hardware moats are slow and structural: actuators are 25-50% of BOM, and ~90% of permanent magnets are made in China. We need to solve supply chain issues. What we underestimate: simulation, non-visual sensing, and safety as its own software category. The next 100X, in my view: omni-models — robots that take in force, acoustic, tactile and chemical data and tell us in natural language what they sense – replacing judgment, not just hands. AI RUNS ON POWER Panelist Jim Messina's numbers were sobering: opposition to data centers has gone from ~45% to 71% in months, 240 cities have moratoriums, and 10 of the 15 biggest congressional races are running ads on them. Tarun Raisoni's counterweight: someone in a lab at IIT or MIT or Berkeley will figure out how to train the model on 1/100 the power. Both can be true — and the industrial base we build for data centers will serve far more than AI. Thanks 🌱🤝🌍 Nicolas Sauvage, Qianran (Katherine) He, PhD, David Delfassy and Starry Wang for another great 100X. 🌱🤝🌍 #Robotics #PhysicalAI #AI #100X

71

2 Comments

Standing here today, with memory chips emerging as a real bottleneck, the “tale of two AIs” letter from David Cahn from Sequoia Capital starts to make sense. Constraints in memory and infrastructure don’t just slow data centers, they create a domino effect on AGI timelines. Progress at the frontier becomes uneven, capital-intensive, and harder to predict. At the same time, watch what’s happening elsewhere. China is pushing LLM evolution along a different axis: model optimization under infrastructure constraints. Less brute force. More efficiency. More architectural creativity. Two philosophies. Keep your ears open. More to come. And learn the new dynamics of “rapidly scaling AI” growth! In the meantime, focus on what matters most, keep building what actually builds the muscle, organizational capability to embrace AI: “AI models are rented. Capability is built.” The HITEC Foundation HITEC Angeles Investors The Tech Series

4

The AI Chip Power Struggle: Who Controls the Future of Compute?  Just gave the following onboarding readiness notes to someone getting into director level role in ASIC AI space. The AI chip market looks chaotic, but in reality, it’s highly concentrated. A few hyperscalers, merchant vendors, and specialized startups control is all we have here.. 1. In 2026, AI chips are no longer just chip with flops and interconnect speed; they are infrastructure platforms. Success depends on compute, networking, cooling, and debugging tool maturaity. Nvidia exemplifies this with systems like the DGX SuperPOD. 2. Despite widespread AI adoption, global compute demand is dominated by a handful of hyperscalers: Google, Amazon, Microsoft, Meta Platforms, OCP and ByteDance. These companies operate the largest AI clusters and are driving 80+% of demand… that is it... if they stop then party stops. 3. GPUs remain dominant not because of raw performance, but because of software ecosystem lock-in. For example, CUDA powers training, deployment, and distributed workloads, creating high switching costs in past - and even with transformer (where its little easy to catch up on new architecture/operators) the close integration of chip to workload where you have large context can be better processed with Nvidia inference context memory storage vs expensive HBM in TPUv7. 4. The strategic divide is clear: hyperscalers build custom silicon to optimize internal workloads (Microsoft Maia , Amazon Trainium), while merchant vendors like Nvidia and AMD sell broadly compatible platforms. And both will have business and growth although some(TPU) will try to enter merchant business - but- it will be very hard for them.... 5. Custom accelerators only pay off at hyperscale utilization, making them suitable mainly for hyperscalers or large AI labs such as Anthropic or AWS. This means no one can just build out on one chip or one large customer demand and this will be a heterogeneous market ( good for neo cloud). 6. Modern AI clusters that are under pressure to scale fast from workloads of agents (clawbots/ claude code) face bottlenecks beyond chips: HBM memory, networking, cooling and datacenter power. Suppliers like TSMC, SK Hynix, and Samsung Electronics are critical in short run. 7. Startups like Graphcore, Cerebras Systems, and Groq can innovate rapidly, but ecosystem lock-in and software integration remain GTM blockers. The clearest opportunity lies in inference, where power, latency, and throughput dominate but that also has not played out for them and will not play well in near future as well… In the end, AI chip market will test the market patience, AI appetite and adaptability- this space will remain in high attention - in a world where attention is all you need. Hope it helps..

29

Zero plus zero equals two. Not in some toy project. In Berkeley SoftFloat -- the IEEE 754 math library reference implementation. It's inside QEMU and most x86 emulators, and it's the oracle that hardware teams verify chip designs against. If you've ever emulated a CPU or validated a chip design in the last decade, Berkeley SoftFloat did the floating-point math. I found seven wrong results across five of six arithmetic operations in the 80-bit extended precision format. The one Intel invented for the x87 FPU in 1980. Some other fun highlights: - 0.5 + 0.5 = 3 - A huge finite number plus zero = infinity - Infinity x 0 = infinity (and no error raised) - Two tiny numbers added together = zero These aren't rounding errors. These are completely wrong answers for valid inputs. The root cause: the x87's 80-bit format has an explicit "integer bit" that every other IEEE format hides. This creates encodings -- unnormals, pseudo-denormals, pseudo-infinities -- where the bit says one thing and the exponent says another. The original 8087 handled all of them correctly. SoftFloat hasn't since its 2011 rewrite. Nobody noticed because the test suite has a structural blind spot. The test generator only produces the encodings that SoftFloat itself would output... the "nice" ones where the integer bit is consistent with the exponent. It never generates the inputs that trigger the bugs. I patched the generator to cover the full input space and failures lit up everywhere. I never would have found any of this if I hadn't been writing my own floating-point library from scratch. When my results disagreed with SoftFloat, I assumed I was wrong. Over and over. I'd go back to my code, recheck my math, trace through my logic... because the reference implementation couldn't possibly be wrong. That's what "reference" means. But the reference was wrong, and I wasn't, and suddenly... Suddenly I was very sad. SoftFloat is supposed to be "the thing that is correct." It's the ultimate tech industry oracle, the final reference on one plus one. TestFloat tests hardware against SoftFloat. FPGA developers validate against SoftFloat. When your personal deity lies, when addition and subtraction themselves dissemble, the epistemological foundation shifts under you. You can't trust the thing you trusted, and now you have to ask what else you can't trust. What makes my situation lonelier is that finding the bug doesn't feel like a win, because it shouldn't have been there in the first place. I wasn't looking for SoftFloat bugs. No one gives you an award for breaking addition and subtraction. I was trying to validate my own work and the ground moved, and now I just feel like I'm waiting for the next earthquake. https://lnkd.in/g67ZCkMu https://lnkd.in/gG7DC-26 https://lnkd.in/g9rJs6ej https://lnkd.in/gV28f9vp https://lnkd.in/g3EPE_wW

7

3 Comments