$PRISM-ML · Prism ML's Bonsai 27B achieves ternary quantization running GPT-3.5 class model on smartphone at 6GB

episode briefing

extracted intelligence

Peter H. Diamandis

China's Kimi K3 Triggers an AI Sputnik Moment, 2.8 Trillion Parameters Go Open Source | Ep. 272

now viewing▶ 63:03

Prism ML's Bonsai 27B achieves ternary quantization running GPT-3.5 class model on smartphone at 6GB

2026-07-19 · 24 company · 22 thematic

sentiment

18 bull5 bear1 neu

speakers

imad

Moonshots panelist, author of 'The Last Economy'. Deep in quantization research (sub-1-bit), edge AI, and inference optimization. Invests in companies working on ternary/binary compression and photonic substrates. Tracks Chinese lab innovations closely.

salim ismail

Salim Ismail is a serial entrepreneur and strategist who co-wrote Exponential Organizations (ExO 1.0 and 2.0). He leads the OpenExO community of 50,000 people in 150 countries and has worked with Fortune 500 companies (P&G, Siemens Energy, HP, Black & Decker) on edge innovation. He is launching the Organizational Singularity (ExO 3.0) as a living AI book via a Claude skill.

dave blundin

Dave Blundin (DB2) is Peter Diamandis' longtime friend and MIT roommate, partner at Link Ventures, and described as a guru on exponential investing. He focuses on moonshot-stage companies in AI, robotics, and space.

immad mostaque

Author of 'The Last Economy'. Tracks Chinese AI labs, quantization, humanoid robotics, and forecasting. Deep knowledge of model architectures, inference providers, and open-source ecosystem. Releasing research papers on generative AI economics.

alex wissner-gross

Computer scientist, physicist, entrepreneur and investor whose work spans artificial intelligence and complex systems. He is the founder and managing partner of Reified.

peter diamandis

Peter Diamandis hosts the Moonshots podcast with a panel of 'mates.' He founded XPRIZE and Singularity University, sits on boards of Trust & Will and Vesmark, and runs the Future Vision XPRIZE for positive AI futures. He frames discussions around the 'singularity economy' and exponential technologies.

quote

“Prism ML and actually Tencent... have had massive advances in being able to take a model that's been trained in a 16bit architecture or an 8-bit architecture like Kimmy is basically 8 bit, 4bit, and take it down to ternary, which is three…”

—immad mostaque

now playing · $PRISM-ML

$FIREWORKS-AIbullish· medium· posimad

Imad flags inference providers as key beneficiaries of open-source model optimization

Inference platforms (Fireworks $17B, Modal $10B, Base10 $10B) have raised massive capital to optimize open-weight models like Kimi K3 for Western hardware, capturing value as ente…

$MODALbullish· medium· posimad

Inference provider Modal at $10B valuation to optimize open-source models

Same thesis as Fireworks: Modal and peers will spend raised capital optimizing Chinese open-weight models for Nvidia/AMD hardware, creating a low-cost inference layer that undercu…

$BASE10bullish· medium· posimad

Base10 at $10B joins inference arms race for open-weight models

Base10 is part of the cohort of well-funded inference providers that will aggressively optimize open-source models (Kimi K3, Inkling) for enterprise self-hosting, accelerating the…

$ANTHROPICbearish· highsalim ismail

Frontier lab valuations could drop 75% as open models commoditize the intelligence layer

Government review requirements cut frontier lab values by 50%, and open-weight models like Kimi K3 matching frontier performance cuts another 50%. The actual frontier cutting edge…

$OPENAIbearish· highsalim ismail

Frontier lab valuations face 75% haircut from open-source competition and regulation

Same thesis as Anthropic: open-weight models (Kimi K3, Inkling) and government review requirements will collapse the premium for closed frontier models, reducing OpenAI's implied…

$ANTHROPICbearish· highsalim ismail

Salem values frontier labs at 25% of prior worth amid open-source commoditization

Open-source models like Kimi K3 and regulatory constraints (FINRA-like oversight) will slash frontier lab valuations by ~75%, leaving only reliability, security, and integrated to…

$THINKING-MACHINESbullish· mediumdave blundin

Mira Murati's Inkling validates open-source self-hosting thesis for enterprises

Murati's departure from OpenAI to build an open-source lab focused on corporate fine-tuning signals that frontier insiders believe self-hosted, fine-tuned open models can match cl…

$NVDA···bullish· highdave blundin

Dave sees silicon demand accelerating despite model efficiency gains

Jevons paradox applies: cheaper, more efficient models (quantization, distillation) will explode total compute demand, keeping Nvidia's high-end chips (Blackwell, Rubin) sold out…

$NVDA···bullish· highdave blundin

Nvidia demand to surge despite Chinese efficiency gains as Jevons paradox drives compute consumption

Export controls forced Chinese labs to develop 100x software efficiencies, but this increases total compute demand rather than reducing it. US inference providers will run Chinese…

$PRISM-MLbullish· highimmad mostaque

Prism ML's Bonsai 27B achieves ternary quantization running GPT-3.5 class model on smartphone at 6GB

Breakthrough in quantization: 27B parameter model compressed to ternary (1.58 bits) with only 5% accuracy drop, enabling frontier-class intelligence on consumer devices without in…

$SSNLF···bullish· mediumimad

Samsung breaks sub-1-bit quantization barrier for on-device frontier models

Samsung's 'nano quant' model achieves sub-1-bit-per-weight quantization, enabling frontier-class models on smartphones. This positions Samsung as a key enabler of edge AI and a po…

$005930.KS···bullish· mediumalex wissner-gross

Samsung publishes nano-quant breaking 1-bit-per-weight barrier for edge AI deployment

Samsung research achieves sub-1-bit quantization via sparsity and low-rank factorization, driven by need to run frontier models on Galaxy smartphones. This will go mainstream with…

$CASSIE-AIbullish· mediumalex wissner-gross

British startup Cassie achieves #1 AI superforecaster, founded by ex-intelligence officer

Cassie (short for Cassandra) tops ForecastBench, statistically indistinguishable from human superforecasters. Founded by British intelligence officer who served in Afghanistan. Hy…

$UNITREEbullish· mediumimmad mostaque

Unitree at 11,000 humanoids produced, scaling to millions as China makes robotics national priority

China has 150 humanoid robot companies with Unitree leading at 11k units produced. MMA combat demonstrations stress-test balance, impact recovery, and latency. Household deploymen…

$OPENAIbearish· highalex wissner-gross

OpenAI valuation under same 75% pressure as Anthropic from open-source commoditization

Same dynamics as Anthropic: government review delays and open-weight frontier models destroy the moat. OpenAI has retreated from owning data centers (Stargate rebranded to leasing…

$SPCX···bullish· mediumalex wissner-gross

Orbital data centers become credible compute layer per SpaceX-Anthropic deal

SpaceX's Starlink/Starship infrastructure enables orbital data centers; Anthropic's collaboration with SpaceX AI for Colossus/Colossus 2 compute validates the thesis. Unit economi…

$SPCX···bullish· highdave blundin

SpaceX Starship rapid reuse demonstrated; orbital data centers inevitable by early 2030s

Starship abort at T-0 showcased extraordinary engineering control. SpaceX AI (xAI) partnered with Anthropic for Colossus 2 orbital compute. Unit economics for orbital vs terrestri…

$FIREWORKS-AIbullish· mediumimmad mostaque

Fireworks raises at $17B to optimize open-source model inference, capturing Chinese model efficiency gains

Inference providers are raising billions to optimize open-weight models like Kimi K3 for Nvidia hardware, creating a 10-100x cost advantage over Chinese labs running on domestic c…

$MODALbullish· mediumimmad mostaque

Modal at $10B valuation building inference infrastructure for open-weight models

Part of the new inference provider layer raising capital to optimize open-source models on Western hardware, capturing the margin between Chinese model innovation and US compute a…

$BASE10bullish· mediumimmad mostaque

Base10 at $10B valuation joining inference provider race for open-source AI

Third major inference provider raising capital to optimize open-weight models, indicating a new infrastructure layer capturing value from model commoditization.

$GOOGL···bearish· mediumalex wissner-gross

Google MIA on frontier cost-performance frontier; may pivot to hyperscaler for other labs

No top Google model on the cost frontier for capabilities. Google should become a hyperscaler provider to other frontier labs (like Meta exploring $10B compute sale to Anthropic)…

$META···bullish· mediumalex wissner-gross

Meta exploring $10B compute sale to Anthropic, pivoting to infrastructure provider

Meta differentiating by going downstack: offering compute to more competitive frontier labs (Anthropic, potentially Chinese models in self-hosting). Also offers Llama models upsta…

$GOOGL···neutral· mediumalex wissner-gross

Alex argues Google should pivot to hyperscaler-for-frontier-labs model

Google appears absent from the frontier cost-performance curve; its strategic move is to monetize its cloud infrastructure by becoming a compute provider to more competitive front…

$MOONSHOT-AIbullish· highpeter diamandis

Kimi K3 2.8T parameter open model triggers Sputnik moment, tops cost-performance frontier

Chinese lab Moonshot released Kimi K3, a 2.8T parameter multimodal open-weight model that jumped to #3 on the cost-performance frontier despite US chip embargo, proving transforme…