In short, Anthropic's Opus 5.5 is insanely impressive. GPT-6 led for a few days but was quickly overshadowed; the pace of progress is too fast and the competition too fierce. The gap between closed-source and open-source models has widened again. This is already the Nth time this year, it's absurd.$Alphabet(GOOGL.US)$Taiwan Semiconductor(TSM.US)
morning_wind
M
This month alone has consumed tens of billions of tokens for frontier models—truly crazy. If we can crack context and verifiers, there's nothing AI can't handle; if you insist otherwise, it just means it's not fast enough yet... $Taiwan Semiconductor(TSM.US)
M
Add more $Alphabet(GOOGL.US); according to various news from North America, Gemini 4 will see significant improvements (compared to 3.x).
M
Continuing to roll selling calls...$Intel(INTC.US)
M
$SoftBank(SFTBY.US)$Microsoft(MSFT.US) tested all day, couldn't stop. ChatGPT-6 is incredibly strong this time around; the key isn't coding, but its extremely fluid compute use and superior long-task performance unlocking new use cases that Claude struggles with. Let's see how these two tickers perform next week.
M
Continue selling hardware calls. $Taiwan Semiconductor(TSM.US)
M![Gemini Gemini Claude Claude GPT-5.6 GPT-5.6 3.8 Flash 3.7Flash Opus 5 Sonnet 5 S]()
$Alphabet(GOOGL.US) released a new 3.8 flash last night, just a few days after the follow-up to 3.7. This is the first officially launched product where RSI (AI self-improvement and iterative enhancement) accounts for a significant portion.
M
$Alphabet(GOOGL.US)$Taiwan Semiconductor(TSM.US) continuing to discuss AI progress and impressions:
(1) LLMs have continued to develop rapidly over the past quarter. On one hand, we see technology diffusing quickly; DeepSeek no longer seems as dazzling. Domestic models like Hunyuan, K3, and GLM 5.3 have made significant progress, giving the impression that they've firmly crossed the kill line and secured their spot on the boat. Meanwhile, OpenAI and Anthropic continue to make substantial strides in coding and beyond with Fable 5.1 this week and the imminent release of 6.0. From the frontier, both OpenAI and Anthropic, along with China's Zhipu, are advancing RSI (where an increasing proportion of AI drives its own evolution, maintaining acceleration).
(2) The main issue with LLMs currently is that coding penetration is very high, essentially a solved problem. Current AI advancements are primarily driven by strong reliability and utility in other domains. However, I suspect it will take time to incubate the scenarios and users who can truly find valuable problems to solve with AI. Among those around me who max out their subscription plans and still complain about not having enough, they are mostly programmers. Other types of users seem unable to unlock the latest AI advancements, due to both the need for exploration and adaptation, and the gradual alignment required in workflow organization and environments.
(3) In capital markets, I believe the poor market performance since August is mainly due to significant uncertainty in the US dollar macro environment and tightening liquidity. Stocks with high elasticity earlier are now experiencing greater volatility, affecting holding experience. I don't know how this will play out or if major issues will arise. But I believe that once the current situation passes, everything will return.
M
$Alphabet(GOOGL.US)$Amazon(AMZN.US) gradually buy back or add some.
M
Let's talk about domestic large models: With the release of kimi3 and the subsequent pullback of deepseek4 pro, it might be time to adjust our rankings of domestic LLM teams. Additionally, $Z.AI(02513.HK) is still very strong, roughly on par with kimi3. I bought a small position for 1000 yuan. I bet that the 'Niu Lai' large model, which has received excellent anonymous test feedback over the past two days, belongs to Zhipu.
M
Let's see who it will be if AI hardware hits a new high in this round?
M
Re-acquired some $Meta Platforms(META.US). After reducing positions in $Alphabet(GOOGL.US), holding steady. Overweight position in $Taiwan Semiconductor(TSM.US) remains unchanged; memory sector has suffered significant losses this round, so I'm ignoring it. For hardware, selling new high calls on sharp rallies; current assessment is that hardware is unlikely to return to June highs. Focus for H2 is on major cloud providers; both Meta and Google show desire to increase positions. Main thoughts and assessments: The narrative of commoditization of models is strengthening, and the competitive-cooperative relationship between cloud and frontier models has entered a stage of mutually beneficial achievement. Overall, cloud may have the upper hand.
M
The calls from a while back were basically 100% collected. Selling more calls today to lower the cost basis.
M
$Alphabet(GOOGL.US) Google's Capex has once again received negative short-term feedback from the market, with investors simultaneously worrying about two things: 1) The portion of Capex invested in model R&D appears to have a low ROI, given the underwhelming performance of Gemini 3.5; 2) The Capex related to selling cloud computing power is suspected of circular transactions with Anthropic, where both profits and backlogs originate from them, as do investments. Double-counting is not allowed, and ROI needs to be discounted. Both of these short-term logics make sense; looking at it over a longer horizon, it might not be that bad: 1) The competition for models may see shifts in dominance, and perhaps the next version will make a comeback; 2) If Liang Sheng's judgment is accurate that "lower cost is ultimately more important," then Gemini being temporarily behind becomes less significant. Cloud infrastructure will be more crucial and grant greater influence, alleviating the above two concerns. The worst-case scenario remains: winner-takes-all in the model space & Gemini continues to underperform.
M
Recent complete sharing by Liang Wenfeng of DeepSeek
1. When we first started this company, our original intention was not to think about how much money we would make in the end, or to go to the capital market, or to list on the stock exchange, or whatever. The initial dozens of people never thought like that; if they had, they wouldn't have joined. 2. We undertook this endeavor with great goodwill towards the world, believing it to be useful to humanity—a matter beyond mere financial gain. Our original intent, our vision, and the vision we have maintained to this day were not driven by a strategy of maximizing commercial interests. 3. Managing a large company...
M
62. The moat of Nvidia's CUDA is being rapidly eroded. On one hand, with the advent of AI, building this ecosystem has become much easier than before because AI can write code. — Liang Sheng from Deepseek.
From this perspective, it is very logical that $AMD(AMD.US) remains largely unscathed in this round. $NVIDIA(NVDA.US) needs a more comprehensive moat, including supply chain, communications, etc.
M
Sell call, sell call
M
1) Cutting-edge models are released here every day, with daily free credits, just like the new energy vehicle launches + new retail subsidy wars; 2) The competition in frontier models is too fierce, the total market size will grow, but how the endgame profits will be distributed is highly uncertain. 3) The current investment in cloud computing here is too large, and it's hard to say how much of the long-term profit share and voice will be taken by the model manufacturers. 4) Hardware truly makes money in the current period. The cost is valuation and discipline.$Taiwan Semiconductor(TSM.US)
M
$Taiwan Semiconductor(TSM.US) Key points from the earnings report and conference call: 1) Gross margin steadily increased; 2) TSMC expects revenue growth in 2026, measured in US dollars, to be slightly above 40%, previously forecasted as above 30%. 3) TSMC expressed confidence in the significant trend of artificial intelligence, stating that capital expenditures over the next three years will be significantly higher than the past three years. 4) TSMC announced that it will raise its capital expenditure ceiling for 2026 to $64 billion.
M
$Taiwan Semiconductor(TSM.US) reported a net profit of NT$706.6 billion in the second quarter, exceeding market expectations of NT$623.73 billion. This is a key compass data point for the hardware sector.
M
$XL2CSOPHYNIX(07709.HK)$SK Hynix(SKHY.US) The US ADR is already at a 50% premium compared to the underlying stock, that's just absurd, absolutely absurd. Some people are shorting the ADR and going long the underlying stock, they're going to have a tough time now.
M
This wave has taken quite a few pullbacks, and many people have been asking for opinions via private messages these past two days. The judgment that AI is a major opportunity on a ten-year scale has not wavered at all. A significant part of the intense pain from this wave of pullbacks is because many people added leverage during the rally, and secondly, it might be due to buying at the peak. If you hold positions with a slightly longer duration, like $Alphabet(GOOGL.US), $AMD(AMD.US), $Taiwan Semiconductor(TSM.US), or DRAM, the returns for this year are still decent, not painful. To summarize the lessons learned, we should focus on these two points, not say AI is over just because the stock price fell. Finally, some believe value will shift downstream in AI, for example, to cloud computing, or to model and application companies (the two are also in a game of their own). I think it will inevitably happen in the long term, but it's not clear enough in the short term at the moment.
