Slowly read it. Researching how tech influences geopolitics. Visiting Fellow, @BayAreaEconomy @dartmouth, Oxford MPhil this fall, first jobs: @CSIS and @CIA
Good morning everyone, some friends and I enjoy circulating our notes on our news. Here's mine for July 18:
Good morning,
> Strikes continue between the US and Iran for the seventh consecutive night since ceasefire talks fell apart. At least 50 people have been killed, according to Iran’s Health Ministry. (CNN)
> Tech stocks had a bad week, chips in particular— with the PHLX Semi index plummeting 10% this week, its steeply weekly drop since April 2025. (WSJ) I try not to read too much into daily swings. Apple edged out Nvidia as the most valuable company again this week as a result, given their patient / passive AI strategy. Seems more noteworthy.
> SpaceX continues to sell around its spare computing capacity. In recent months its been inking deals with Anthropic and Google… and now headlines say the company is in talks for a multibillion+ deal with the DoD for access to its data centers to run high-side AI models. Pentagon already has a relationship with them for rocket launches and satellite management. That’d be cool analytical work to return to one day. (WSJ)
> AI Model analysts continue to raise alarm bells about Kimi K3. Whether it’s this Sputnik moment is something I’ll be tracking to make sense of for my own personal opinion. I’m not really sure. The finance crowd attributes it to more market uncertainty with all the capex in US markets and people wondering if that’s a bright idea. (X)
> Cyclosporiasis, a food-borne illness in lettuce, has spread to more than 30 states. Taco Bell got flack for its lettuce supplier hitting 5 states. There have been more than 1,644 confirmed cases with 5,100 additional suspected. I found a great explainer article that I’d be happy to pass along. (WSJ / CDC)
> The Andes are getting absolutely dumped on— upwards of 200 inches of snow could arrive over the coming days. An enormous blizzard is hitting resorts across South America. Ski Portillo, Chile, is already reporting 41 inches in the past 24 hours. 3.5ft. Wow. (https://t.co/vAYomlKl7Z)
> France vs England today at 2pmPT / 5pmET; Spain v Argentina tomorrow at 12pmPT / 3pm ET. Lots of storylines to track here: the Golden Boot race, Argentina repeat, the weird nature of a 3rd place game, etc. What are you watching for?
@besttrousers College grads are definitely feeling the impact of hiring from AI— or at least those are the claims by the powers at be. I don't think that's disputed. I've seen corporate cycles disrupted, startup offers entirely rescinded, and new roles sprout up and old ones be killed.
@AsiaLens@RahmEmanuel Hey Demetri, I've been trying to follow Aspen along from afar, is there anyone there doing some solid inside baseball live reporting? Providing commentary over that panels? Thanks
I did some quick tests that indicated that the Kimi K3 pretrain is around halfway between Opus 4 and Opus 4.5. So ~10 months behind Anthropic. These tests probably understate data improvements, so overall I think it's a similarly good pretrain to Opus 4.5 (~8 months behind). These tests are better at measuring "general pretrain capability" than at incorporating (coding-specific) data quality.
This was prompted by me thinking more and realizing that my claim that "As a pretrain, it's probably somewhere between 4.8 and Mythos (around halfway between?)" was probably too bullish on the model and that I might as well test and find out. (And yep, this was very wrong.)
I think Mythos is a pretty big step up in pretraining, so K3 might be more than 8 months behind on the historical pretraining trend relative to Mythos (as in, Mythos is >>3 months ahead of K3 and Mythos was fully done training ~5 months ago).
Overall, this makes me suspect more of the improvements are due to distillation-type effects and makes me think the full catch-up times would be somewhat longer (if Ant/OpenAI stopped but investment still followed current trends). Minimally, more of the improvement probably lives in post-training/mid-training.
For reference, the same test indicates K2.6 is around halfway between Sonnet 4 and Sonnet 4.5. (And this roughly corresponds to some other similar measures.)
Sorry about the error.
Hey everyone, my friends and I like to read our news with some intention every morning by trading notes on it. Here's mine for Friday, July 17:
Good morning,
> See those hazy pictures everywhere online? That’s the result of wildfires from Canada pouring into major cities across the Great Lakes, Northeast, and Mid-Atlantic. AQI is terrible for over 100 million people through Saturday until the wind, rain, and storms are expected to provide some relief. (CNN Weather)
> Trump gave a primetime address from the White House. He discussed election interference and declassified new evidence of vulnerabilities in our system. I don’t really see much unique signal emanating from this speech versus his tendency to blast the airways with questionable claims in, well, just about everything. He’s just doing more groundwork to question future elections— scary. The speech made headlines across outlets but not leaving his usual choir and audience. Feel free to fight me here. (CNN / WSJ / NYT)
> TSMC raised its outlook in its quarterly earning call. I can link the stats but analysts also note that it’s spending additional capex to diversify chipmaking away from Taiwan, including the US $100B+; we’ll see if that pans out. Fabs are notoriously hard to put elsewhere and create a local Los Alamos-like economy around. (MSS / Tech Cold War Book)
> Some noteworthy news in the US v China AI competition" of frontier models; Moon Shot AI (?) released Kimi K3; it’s up there with Fable 5 and GPT 5.5 in benchmarks and beating them in Code Arena. (Kimi K3) Now you’ve got op-eds galore raising alarm bells about China erasing America’s AI lead. I’m still bullish on the US longterm for being a leading innovator, even if China might have the diffusive edge currently… (Axios)
> China also established a World AI Cooperation Organization (WAICO) with 29 founding members, including Russia, Indonesia, Pakistan, Brazil, South Africa, and Vietnam. Its goal is to promote cooperation on governance standards. (Bloomberg)
> Christopher Nolan’s The Odyssey is killing it on Rotten Tomatoes 96%, the leading aggregator for critics’ reviews. I’ve read a few reviews and can’t wait to watch it a bazillion times.
> Headed to SLO this weekend with friends to spend and spend much time outside as possible and read. Send anything I should catch.
Definitely. And I think as the digital tech industry continues to consolidate and box out new players, it will also mature into a "lobby" (akin to oil, agri, pharma, etc). They might compete in the Bay but will increasingly (openly) cooperate in Washington. This happens at the tippy top already but that realization is going to diffuse more across smaller players.
Each morning, I read the news and trade notes with friends. Here are mine for Thurs July 16:
Good morning,
> Chevron is announcing its intending to sign preliminary deals this Friday to invest in two Iraqi oil fields. The latest blip of noteworthy news as supply chains for oil shift around the Strait. (WSJ)
> Floods in central Texas are turning deadly. I’ve got CNN on and one person has been killed. One of my old friends did search and rescue in previous stormy seasons (including Camp Mystic); I remember her telling me how its this combo of the flash floods + residents’ stubbornness that lead to tragedy striking these communities. (CNN)
> Zelensky dismissed his Defense Minister Mykhalio Fedorov, sparking protests in Kyiv and other Ukrainian cities. The articles I’ve read converge over saying it’s due to internal tensions. Notably in my research on tech companies influencing geopolitics, he was the one that asked SpaceX founder Elon Musk to stop Russia from using Starlink satellites for drone attacks. (BBC)
> DeepSeek is rumored to be planning for an IPO after another fundraising round its got planned at a $74B evaluation. They raised $7.4B in June. I’m throwing these numbers around because it’s significant to show the magnitude of capital with how much money is flowing in USA frontier AI versus China. Deserves more of me blabbing here. (TechCrunch / Reuters)
> Inflation growth slowed to 3.5% in June, beating expectations and improving off of 4.2% in May. I missed this; WSJ analysts suggest it was due to gas prices managing to drop. Good to just always have a grasp on where inflation lies, even if your microeconomy comes first before US macro. (WSJ)
> Thinking Machines, a leading neolab, released their first model “Inkling”. I read a lot about it last night and want to try it. Benchmarks suggest it’s the best American open-weight model, behind your frontier Fable’s and Sol’s, but it’s going to allow easier customizability for enterprise and power-users alike. [Thinking Machines is led by OpenAI’s Former CTO Mira Murati. She went to Dartmouth, wooo!] (MTS Substack, WSJ)
> The final is set between Argentina and Spain this Sunday at 12:00pm PT / 3PM ET. I grew up watching him on Pep’s Barca— needless to say, this is a great final for me. I’m stoked.
> I swam the Alcatraz Escape route this morning before work. I love open water swimming and can’t wait to continue to see where the sport takes me. So many parallels to rowing.
TSMC 2nd quarter (Q2) conference on Thursday (7/16), media report expectations:
Q2 revenue NT$1.27 trillion (already known, US$ not known)
-Net profit seen up 59% to NT$632.6 billion (US$19.6B), 10th straight quarter of growth
-Gross margin could exceed 68%
-Operating margin seen hitting new high
-R&D expense seen hitting record high due to advanced process investments
Q3 revenue forecast seen +9%-12% vs Q2 in USD terms (some 15% estimates)
Full Year
-Raise full year revenue to mid-30s% increase from current ‘above 30%’ growth, most bullish scenario sees 40%
-Unlikely to raise 2026 capex (some see $58B-$60B) but could discuss plan for next 3-years
5-year CAGR forecast 2024-2029:
-AI semiconductor revenue CAGR for 2024–2029 seen raised to 70%-80% from mid-to-high 50% now
TSMC Official Q2 Guidance
-Revenue US$39.0B to $40.2B
-Gross Margin: 65.5% to 67.5%
-Operating Margin: 56.5% to 58.5%
-Exchange rate forecast NT$31.7 per US dollar.
What investors want to know:
-Overall AI demand health
-CoWoS packaging capacity expansion
-N2/N3 ramp and pricing
-Is memory shortage hurting demand for logic chips?
-Update on global construction projects
-Update on advanced processes and packaging (CoPoS, SoIC, etc) R&D
-TSMC view on intensified competition from Intel, Samsung
-Will TSMC raise prices again in 2027? (5%-10% increase expected)
-Gross margin projections for 2026 and 2027
-Capex strategy for next 3-years
-Semiconductor industry outlook
Links:
https://t.co/jRoAHHj7IB
https://t.co/tUdnIEcJ0n
https://t.co/8bA7qCjV4q
https://t.co/HOIoQDCtl7
https://t.co/0GTALbLkiM
https://t.co/WOgHeoAhVA
https://t.co/htUJGlFjSR
https://t.co/ZPOmnZMCRf
https://t.co/bJozxXJDKE