Nicholas Bardy

🚨 Shocking: Frontier LLMs score 85-95% on standard coding benchmarks. We gave them equivalent problems in languages they couldn't have memorized. They collapsed to 0-11%. Presenting EsoLang-Bench. Accepted to the Logical Reasoning and ICBINB workshops at ICLR 2026 🧵

Back when we had the 1m context window in Amp, we noticed our numbers go up MASSIVELY. There was a small number of users who did *everything* (think: multiple projects!) in the same thread over multiple days (cache is gone). Crazy costs. We reached out and told them about it.

I'll call my shot more specifically: Q1 2027 an open source model running on 32gb of memory will do 100 tok/s and match Opus 4.6 on most benchmarks

Be the person with $100k in the bank and a $5k car. Not $5k in the bank and a $100k car.

Midjourney v8 alpha is out now. Here are 20 examples using their new native HD, no SREFS, text only. My thoughts @ the end, plus 3 exclusive prompts for subscribers from these 3 below.

The $LITE thesis: The hidden monopoly in the AI. Lumentum is up 316% YTD, but might be 1000%+ by 2027. Micron ($300B) or TSM ($1.5T) sit in the center of every TPU/GPU deployed. But same with $LITE, but it's a $26B MC. In Every, Single, TPU from Google, $LITE makes unbelievable amounts of profit for their marketcap. That's because it's the standard for Optical Circuit Switching (OCS) + optical networking. It's also in - $NVDA Blackwell -$AMZN Trainium - and other hyperscaler ASICs. Lumentum sits in the holy trinity of every single chip deployment for photonics. And for every TPU capex spent, $LITE takes 8-12%. For every Nvidia GPU, $LITE takes ~2-3% (split between Innolight and some others, so the math gets a bit complex). But some napkin math on NVDA GPU deployments alone for BOM: NVIDIA Blackwell (GB200): HBM memory: ~50–55% (SK Hynix (Lead), Micron, Samsung) Logic (GPU Die): ~25-30% ( $TSM 4NP) CoWoS Packaging: ~13-18% $TSM Optics/Network: ~3–5% (Innolight, Lumentum, Coherent) PCB/Power: 5% For Google TPIU "Ironwood" TPU v7: HBM Memory: 38-42% Samsung / SK Hynix Logic Die: TSM ~28-33% Design/I.O: 8-10% MediaTek Optical Network: 10-14% ( $LITE (primary), $COHR secondary) Optical Switch: 2-4% $LITE $LITE est. total cluster share: ~8–12% Just an FYI, Google's "Optical" BOM share (8–12%) is an anomaly due to their unique Optical Circuit Switch (OCS) monopoly. Just for some napkin math: $40B Google TPU spend by 2027. $LITE captures 10% (30-40% margins), $1.5B+ FCF from Google alone, 17x earnings from just their primary customer. (analysts are probably extremely off with projecting TPU spend scaling). Not even including their split from $AMZN Trainium, $NVDA Blackwell, $MSFT Maia, and other chip deployments. $LITE is in the center of every single TPU/GPU future chip deployment for now and takes a cut. The only downside is they're the clear market leader now, but $AVGO and $COHR are likely set up to compete by 2027-2028. However... People say "$26B, ATH, why are you buying now". This is the reason. They're involved in every future single TPU/GPU/ASIC deployed. $LITE could end up easily over $60B+ if Google TPUs, and other chip spend ramps up and LITE takes a 2-3% (from $NVDA, $AMZN, $MSFT) or 8-12% cut (from $GOOGL) for every single dollar spent.