🕳️ Mole for Mac 1.12 is live. The theme this time: use less of your Mac, and be honest about what it did. Lighter over days, not minutes. Sampling wakes less often, menu bar CPU stops creeping up, long lists stay smooth. Cleanup reaches further: Dia, Zed, stale caches. It marks cloud-synced files, reports the space you actually got back, and never stalls on a stubborn folder. Chat history, local models and active builds stay untouched. Updates that finish. App Store apps update inside Mole when that's safe, Homebrew shows real progress and survives slow networks, ChatGPT and Chrome hand off to their own updater. When macOS blocks a replacement, the row tells you why and opens the right setting. Battery Care can top up to 100% on demand, and the card shows live power draw, real health, and whatever is eating the most. The menu bar runners are redrawn pixel-crisp from the open-source RunCat Neo gallery (thank you, Takuto Nakamura), and every runner's pace tracks real load. Plus: Celsius by default, an Open Source Licenses panel, SMART failures in the health score, VoiceOver fixes, and no more twin windows on dual-display Macs.

Re: High-NA Chips in Production. This means an alternate runpath to the POR multipatterning process. It's common to have competing integration flows for some modules. It allows the fabs to directly compare yield, cost, and cycle time for the new process. For memory, they may even allocate a portion of the WIP to the competing flow and sell it if it yields, although that's rare.

Tesla ordering parts for 15000 Optimus start September through December and scaling to 1000 per week by end of September 2026 and 2000-2500 per week rate by end of December 2026. 125,000 runrate for 2027. $TSLA Congratulations to the Tesla Team @Tesla_Optimus @Tesla_AI @elonmusk

Imagine Celestial ($MRVL acquisition) was public. The valuation would probably be astronomical in US markets (if I had to guess, $6-10B) Because 2028 revenue projections were $500m (q4 annualized runrate), 2029 revenue projections were $1B. And because Celestial is a core part of hyperscaler CPO programs. Celestial's 2026 Q2 earnings would have been: - Close to $0 revenue (Marvell said "revenue and earnings since the acquisition were not material") - Losing -$12.5M/quarter ($50M/year) Would people who knew what they were talking about looking at qualification cycle players say... 1. based on Celestial's Q2 P/S numbers and because it was losing -$12.5M a quarter... That it's a worthless meme stock? Or 2. with CPO hyperscaler opportunities for 2028, it should be valued at $6-10B? And I think that's a core cultural disagreement between US / EU cultures + markets. If you knew how to evaluate qualification cycle players, which number actually matters? I'd argue overwhelmingly the latter.

Pocket FM, an audio drama platform, is a fascinating example of how AI can transform content business. Content production is now 80x cheaper, and output is up from ~100K hrs/yr to 2.5M hrs/yr. As a result, retention is up from 44% to 76% and revenue runrate doubled to $500M

I feel like traditional "value investing" went into hibernation. Before it was finding a 10x -12x P/E company like $UPWK or $PYPL, Dominoes, Pepsi. Or even a cheap Japanese company below book value. Then just enjoying slow growth but buyback/dividend appreciation from their cashflow. Now you can find 1.8x runrate P/E company like $ESMT that grew revenue 605.24% Y/Y. And the main thing you need to argue about is duration. Or a 16.2x trailing P/E Sumitomo, despite having way too many hyperscaler LTAs in AI from fiber to lasers. AI created much bigger medium-term anomaly, where "value investing" went towards high growth, "deep value" companies.

I wanted exposure to DDR2/DDR3 bottleneck ongoing and found ESMT (3006). A $2.5B MC fabless company with PSMC wafer allocation (esp. focused on DDR2). Last month's net income was: $109.5M (July month), which annualized is $1.31B net profit (1.9x runrate P/E). If we look at net profit throughout the months tracking legacy DRAM price hikes: Jan 2026: ~$16M (est) Feb 2026: ~$17M Mar 2026: ~$31M (est) April 2026: ~$58M May 2026: ~$60M (est) June 2026: ~$67M (est) Jul 2026: ~$110M (Estimated months are inferred from reported quarterly totals. July is company reported) This earnings progression reminds me of $SNDK style price hikes + earnings inflection (esp. the report released this month). And I do expect DDR2/DDR3 capacity to remain constrained throughout 2027 (with legacy players like Winbond withdrawing from certain lines too like DDR2). It's not quite all inventory liquidation like the other peers. Sure lower cost inventory helped, but the primary driver seems to be the widening spread between wafer costs/supply and legacy dram ASPs? Their balance sheet is extremely solid as well: - Cash on hand: ~$395.9M net cash - $249.6M inventory, $146.3M receivables And the net income progression... is before further legacy DRAM hikes expected in Q3. *disclosure own positions, NFA It's always a bit daunting being early without much commentary around the idea. Wondering if anyone can stress test this thesis? Since a company that grew monthly net income from $16M -> $100m this year alone (July Annualized would be 1.9x P/E). And benefits from another projected wave of DDR2/DDR3 hikes this quarter... maybe like 35-40% for DDR2 lines per Trendforce, and DDR3 likely continuing to rise. Do markets just not know about this bottleneck/company or am I missing something from my research?

Today, @proximalhq is announcing its $15M seed round led by @generalcatalyst, and @Chemistry is proud to be part of the round. Some founding teams you meet and immediately start rooting for. Proximal is one of those. We back people first, and this team made that easy. @calvinchen had already built and sold a company, and continues to be one of the highest velocity builders we know. @MatternJustus led RL research at Prime Intellect and is well recognized as one of the experts in the space. But what stayed with us was how they work together. They brought research depth, systems chops and product instinct to one founding team, and they moved so fast that the bar went up every time we met. Proximal builds systems that automatically evaluate AI models, find where they fall short, and create the post-training data to fix them. We think closing that loop is one of the last unlocks for models that recursively self-improve, for faster progress across science and engineering, and for enterprises actually putting AI to work. And they're just getting started. Calvin, Justus: we're thrilled to be on this journey with you since day one. At $200M runrate revenue, it's been an incredible 10 months.