Reading

Taiwan-based semiconductor company MediaTek announced at its October 30 earnings call that its first customized AI accelerator (ASIC) for data centers began shipping initial orders to customers in September. This marks MediaTek's formal expansion from mobile chips into the high-growth AI infrastructure market.

Driven by customer demand, MediaTek revised its fourth-quarter cloud ASIC revenue estimate upward to $2.5 billion. For 2027, the company stated it aims to reach the upper end of its original target range—$16 billion—which would represent approximately 20% of the global AI ASIC market. Additionally, MediaTek confirmed securing a second cloud ASIC customer, with related projects expected to contribute revenue starting in 2028.

Despite the optimistic outlook, MediaTek CEO Rick Tsai noted that converting orders into actual revenue depends on whether the manufacturing supply chain can keep up with demand, particularly regarding substrate supply challenges. Therefore, the $16 billion figure remains a next-year revenue target rather than realized shipment performance. The company is working with suppliers to secure more capacity and advancing next-generation chip designs to improve compute efficiency per watt.

Researchers at Adobe Research (US) and UC San Diego (US) have introduced EditAgent, an AI image editing system designed to address the high failure rates of general-purpose editors when handling complex instructions. The core mechanism involves breaking editing tasks into sub-steps and generating multiple candidates in parallel by pairing a general-purpose editor with specialized tools for operations like sharpening, rotation, and zooming.

Unlike traditional methods that rely on retrying the same editor, EditAgent employs an unmodified standard language model as a judge. This model receives verification records containing close-ups of edited areas, change maps, and measured statistics on brightness and color. The study indicates that these hard numbers help the standard model identify errors such as overexposure or color temperature shifts caused by the general editor, allowing it to select superior results.

In a benchmark test of 600 image edits, blind-rated by five professionals, the results showed that when provided with verification data, all seven tested standard language models, including Qwen3.8-27B, outperformed the best fine-tuned evaluation model, EditScore-72B. Specifically, Qwen3.8-27B achieved an 81.3 percent accuracy rate in selecting usable results, compared to 74.8 percent for EditScore-72B. This suggests that the structured input data may matter more than the training method for evaluation models.

Although EditAgent improved the performance of open-source editors Qwen-Image-Edit-2511 and Flux-2-Klein-9B across three benchmarks and outperformed seven other AI agents, its primary limitation is computational resource consumption. The agent requires approximately ten times the compute power of a single editor operation. Additionally, researchers used about 26,000 verified training examples generated by EditAgent to fine-tune editors, resulting in performance improvements, which implies that the high compute cost might be partially offset by the value of the generated training data.

In the latest episode of the Equity podcast, TechCrunch's Amanda Silberling offered an observation: current AI executives' product narratives are not merely marketing blunders, but a hybrid of product and marketing issues. She argued that the tech industry is pushing toward a future where basic human interactions are increasingly optimized out of existence, adding that humans being able to talk to each other is likely a net positive for society.

Silberling pointed out that the root of the problem lies in executives at these extremely high-value companies sitting together discussing pain points in their own lives, focusing on scenarios like booking dinner reservations or flights—situations that do not constitute real challenges for ordinary users. She noted that while everyone books flights, she does not do so once a week, nor does she frequently book dinners; even when she does, they are not particularly difficult tasks in her life. She believes tools like ChatGPT and Claude do have mass appeal, but the value ordinary people derive from them may not align with the narrow scenarios executives describe.

TechCrunch weekend editor Anthony Ha reviewed a recurring pattern in the tech industry during the discussion. He recalled being confident that Uber would never gain broader traction because it initially targeted San Francisco's early tech adopters—a narrow audience of young urban millennials with disposable income for cars. Ha noted that the tech industry has long succeeded by betting that people have not yet caught up with the frontier, and that frontier companies represent what future consumers and families will look like. In the case of AI, however, this bet seems unproven, partly because the society into which the technology is being promoted is very different.

Silberling observed a parallel phenomenon: alongside over-optimization in the tech world, there is a movement of people reverting to flip phones or buying physical CDs, tired of algorithms telling them what they want to hear. She sees this as reflecting a cultural tension: people have not yet found the balance between what optimization truly helps them and what merely makes them feel less human than they wish to be.

The market context for this discussion is Anthropic's preparation for an initial public offering in the United States, as Wall Street begins losing its appetite for heavy AI spending. If AI companies' core narrative—that making everyone's life easier—fails to resonate with ordinary users, their growth story will face stricter scrutiny in public markets than it did in private stages.

However, this argument has its boundaries. Executives' public statements may be marketing rhetoric aimed at investors rather than an actual reflection of product roadmaps. Silberling herself acknowledged that people clearly find value in ChatGPT and Claude. The issue is more likely that the narrative extends too far into scenarios like booking flights and dating, rather than the products themselves being entirely useless to average users.

Datacenter construction in the United States for 2027 is largely locked in. In an analysis released on November 1, SemiAnalysis estimated that roughly three-quarters of the 33.8 GW of delivery capacity across its top 20 states scheduled for 2027 are already under construction, meaning next week's midterm elections cannot prevent their completion. The real battleground lies in 2028: the model predicts a net addition of 62.2 GW in the US that year, most of which has not yet started construction, and these unstarted projects are highly concentrated in states electing governors on November 3.

The single largest current risk is the permit freeze at the Texas Commission on Environmental Quality (TCEQ). Texas accounts for nearly 30% of the new US capacity added in 2028 within the SemiAnalysis model. Since September 21, TCEQ has frozen new permits, and combined with the interconnection queue audit begun on August 3, this pause now also covers onsite power generation projects. Two scenarios cited by SemiAnalysis differ significantly: Republican incumbent Governor Greg Abbott may resume permit issuance after the Electric Reliability Council of Texas (ERCOT) submits its audit report on December 10, resulting in an interruption of about three months; Democratic challenger Gina Hinojosa advocates maintaining the freeze throughout the legislative session (January 12 to May 31, 2027) until new laws are passed, meaning the pause would last about eight months from September. If developers must wait further for new rules to take effect in late summer, the interruption approaches one year—five to nine months longer than a December restart. For campuses racing toward delivery in 2027, this time window may be untenable.

SemiAnalysis divides the 13 states with the largest 2028 net additions (accounting for 70% of the national total) into three risk tiers based on election scenarios. Tier 1 includes Texas, Ohio, Arizona, and Georgia, where unstarted capacity exceeds 2 GW and stricter conditions will apply regardless of who wins—Texas, Ohio, and Arizona have candidates from both parties leaning toward tightening regulations, while Georgia is included due to a tight governor's race and simultaneous elections for majority seats on the Public Service Commission. Tier 2 comprises Pennsylvania and Oklahoma, where additional restrictions depend on whether there is a change in party control. Tier 3 includes Illinois, Oregon, and Iowa, where unstarted capacity is smaller but regulatory tightening is almost certain. If all candidates advocating for a moratorium are elected, approximately 30 GW of planned 2028 capacity across these states faces delay risk; if only candidates advocating for conditional requirements are elected, exposure drops to 5.7 GW.

Bipartisan consensus centers on three points: making datacenters bear their own grid costs (advocated by 15 candidates), eliminating tax breaks (12 candidates), and limiting or disclosing water usage (11 candidates). SemiAnalysis argues the key difference between the 'conditional' and 'moratorium' camps is that conditional demands alter development costs and terms, allowing developers to continue by funding grid upgrades, adjusting cooling designs, or negotiating community agreements. Moratoriums, by contrast, directly block projects that still require necessary permits, preventing them from starting. The article cites interviews with six current and former employees of hyperscale cloud service providers, noting that while these companies can afford higher compliance costs, having money does not guarantee on-time delivery—permits, equipment, and construction schedules remain constrained by freezes, and shifting development to other states is difficult because any new location would similarly require power, permits, and construction arrangements matching the delivery timeline.

Trucks VC, a US-based venture capital firm specializing in early-stage investing, recently published a blog post stating that it has received business plans from 81 new autonomous vehicle (AV) startups so far in 2026. This figure represents a two-thirds increase compared to last year and exceeds the volume seen during the industry's boom period a decade ago.

Puneeth Meruva, a partner at Trucks VC, described this as a "second, even bigger AV wave" fueled by technological breakthroughs. He argued that foundation models now provide a viable technical architecture for end-to-end autonomous systems, enabling cheaper and faster deployment. The availability of powerful open-source vision-language-action models, such as Nvidia's Alpamayo and Qwen, has significantly lowered the barrier to entry.

Meruva noted that many startups are deploying within months on less than $1 million in funding. Beyond traditional applications in construction, warehousing, and mining, an increasing number of startups are pitching pure-play robotaxi businesses, positioning themselves as direct competitors to Waymo. While these smaller players may not compete globally with Waymo, Meruva believes they can find niches in specific regions or route types.

The surface data of the U.S. economy appears robust—real GDP continues to expand at an annualized rate of 3.4%, consumers keep spending, and businesses keep investing. But Invictus, the author behind The Big Picture blog, points out in this analysis that a less conspicuous problem is emerging: the share of recent economic growth coming from information-processing equipment and software is unusually large.

Comparing the five-year expansion period from 1995 to 1999 with the most recent eight quarters ending in September 2026 reveals two very different pictures. In the late 1990s, average annualized real GDP growth was approximately 4.17%; over the latest eight quarters, it was only 2.34%. More noteworthy is the composition of growth: investment in information-processing equipment and software contributed an average of 0.68 percentage points to quarterly annualized growth in the late 1990s, rising to 0.81 percentage points over the most recent eight quarters. In absolute terms, the pull from technology investment is actually stronger now, but the proportion is the issue—in the late 1990s, IT and software accounted for about 16% of total growth, whereas this ratio reached 35% during the period from 2024Q4 to 2026Q3.

Invictus uses a simple thought experiment to highlight this shift: subtracting the contribution of IT and software from total growth. In the late 1990s, the rest of the economy contributed approximately 3.49 percentage points of annualized growth after excluding IT; over the most recent eight quarters, this figure was only 1.53 percentage points, a decline of about 56%. He specifically emphasizes that this does not mean the non-IT economy is contracting, but rather that its contribution to total growth is much smaller than it was in the 1990s.

Other data further corroborates these concerns. Residential investment had a positive contribution to GDP growth in the late 1990s but turned negative over the most recent eight quarters; consumer spending remains an important source of growth, but its contribution is far less significant than before. Invictus uses a metaphor: several legs supporting the chair that is the U.S. economy are wobbling. The economy of the late 1990s had multiple powerful engines running simultaneously, whereas today’s economy relies more heavily on just a few.

Why is AI capital expenditure a core variable? Invictus explains that companies are pouring huge sums into computing power, data centers, networking equipment, and software, anticipating future demand and productivity gains. These expenditures count toward GDP today, but whether they ultimately generate sufficient financial returns is another question. If expected returns are disappointing, businesses may scale back investment—the author believes signs of this have already appeared. A reduction in investment would directly weaken GDP growth and could indirectly affect employment, construction, suppliers, and consumer spending, with economic consequences potentially extending well beyond the technology sector itself.

However, Invictus simultaneously provides important caveats. First, the BEA’s classification covers a scope far broader than artificial intelligence itself; the 35% share does not mean AI contributed 35% of economic growth; these categories capture important components of the entire investment boom surrounding computing infrastructure and software. Second, this counterfactual comparison is not perfect: a significant amount of AI capital expenditure has shifted from other investment areas because it is the current hotspot, and managers tend to chase performance; without the AI investment boom, this capital might have flowed to other industries. Furthermore, the late 1990s were hardly free of speculative bubbles—the dot-com bubble eventually burst, leading to a recession in 2001. Invictus also acknowledges an important distinction between concentrated growth and systemic fragility: American households, banks, and corporations currently may still possess the financial strength to cushion against investment reversals, and U.S. consumer resilience continues to exceed expectations. GDP contribution data alone cannot answer this question.

Global cryptocurrency exchange Binance announced early on Sunday, November 1, 2026, that it would begin allowing withdrawals in a phased process. This follows the freezing of all deposits and withdrawals after hackers stole $6.3 billion worth of crypto from Binance wallets. According to TechCrunch, Binance stated in its latest update that it has now "covered 95% of the stolen assets" and is still working to buy the "remaining affected assets," which are described as less liquid and requiring more time to source at scale.

Binance emphasized: "Our commitment remains unchanged: Binance will bear the financial impact of the stolen funds." The company stated these losses will not be passed on to users. Regarding the initial freeze, Binance noted it was a necessary precaution while teams worked to contain the incident loss, assess impact, restore services, and complete security checks.

Withdrawal resumption depends on network and asset type. As of Sunday morning, Binance's network status page showed mixed statuses: Bitcoin withdrawals and deposits remained partially suspended, while Ethereum deposits were normal but withdrawals were partially suspended. Users were warned to expect temporary delays due to traffic volume.

On attribution, Binance previously blamed North Korean hackers. On Saturday, the U.S. Federal Bureau of Investigation (FBI) confirmed a North Korean group known as TraderTraitor was responsible, stating the group was converting assets primarily into Ethereum with further laundering likely. Binance plans to publish a more in-depth incident report upon completion of its investigation.

In the book Free Capital, private investor Vernon describes a strategy: buying growth stocks when they suffer negative events that cause an overreaction in the market. He calls this a 'glitch'—a temporary market panic rather than a fundamental change to the business model. Asian Value Investing newsletter author Michael Fritzell applies this framework to South Korean e-commerce company Coupang (CPNG US, market cap approx. $26 billion): The company's stock price dropped significantly after its November 2025 cybersecurity incident. Is this a 'glitch' or structural damage?

Fritzell believes Coupang possesses several structural characteristics that may support its recovery. First is the WOW membership system: for a monthly fee of approximately KRW 7,890 (about $5.9), it bundles services such as free delivery, free returns, online fresh groceries, and the video streaming service Coupang Play. This bundling increases the switching costs for users leaving the platform. Second is logistics scale: with approximately 78 million square feet of warehouse facilities and most of its 108,000 employees working in warehouses, the goal is to have nearly the entire population of South Korea living within 10 kilometers of a Coupang logistics center by 2027. Fritzell points out that the more customers there are, the more suppliers can be attracted to its proprietary e-commerce business; the larger the company, the higher the return on its logistics infrastructure.

Another key factor is founder Bom Kim's control. He holds all Class B shares, which carry voting rights 29 times greater than those of Class A shares. Fritzell writes that investing in Coupang is largely a bet on Bom Kim himself and his long-term ambitions. In its business presentations, Coupang emphasizes five principles: customer-centricity, embracing hard work, using technology to solve customer problems, pursuing long-term cash flow growth, and adopting a return-on-capital mindset. Fritzell notes that although the company is still in an investment phase and its financial data remains unclear, it is rare to see a South Korean enterprise so explicitly target free cash flow generation and high capital returns.

However, Fritzell himself does not draw a conclusion. He explicitly states that he is trying to figure out whether this is a glitch or a lasting change, and declares at the beginning of the article that as of October 11, 2026, he does not hold a position in Coupang, receives no compensation from any mentioned companies, and all estimates and forward-looking figures are based on assumptions that may prove wrong. This analysis provides an evaluation framework and structural arguments, rather than a verified judgment of recovery. Whether the cybersecurity incident was a temporary glitch or a lasting change to the business requires subsequent financial reports and operational data to confirm.

Before 2014, mainstream neural machine translation models mostly adopted the encoder-decoder architecture. This architecture compresses the source sentence into a fixed-length vector, which is then used by the decoder to generate the translation. However, when the input sentence is long, this "bottleneck" leads to information loss, making it difficult for the model to handle long sentences. Bahdanau et al. pointed out in their paper that using a fixed-length vector was the main obstacle to improving performance.

To address this issue, the authors proposed a new extended architecture that allows the model to automatically search for relevant parts of the source sentence when predicting target words. Instead of attempting to encode the entire input sentence into a single vector, this method encodes it into a set of annotation vectors. During decoding, the model dynamically selects a subset of these annotation vectors as context vectors based on the currently generated word. This mechanism later became known as the Attention Mechanism.

Experimental results showed that this method achieved performance comparable to traditional phrase-based systems on English-French translation tasks, with improvements being particularly evident in long sentences. More importantly, qualitative analysis indicated that the soft alignments found by the model were highly consistent with human intuition. For example, when processing long sentences containing complex subordinate clauses, the model could accurately map specific vocabulary from the source language to corresponding positions in the target language without explicitly segmenting hard fragments.

For today's readers, understanding this mechanism is crucial because it forms the foundation for subsequent Transformer architectures and large language models. Suppose we are building a real-time translation application; if we still use early fixed-vector methods, long paragraphs entered by users may result in semantic breaks or omissions. With the introduction of the attention mechanism, the model can "focus" on key information like humans do, thereby significantly improving translation coherence and accuracy. This is not just a technological iteration, but a key step in simulating human cognitive processes.

"Neural Machine Translation by Jointly Learning to Align and Translate" was published in 2014 by Dzmitry Bahdanau, KyungHyun Cho, and Yoshua Bengio. This paper is suitable for researchers who wish to deeply understand the underlying logic of natural language processing, AI engineers, and students interested in sequence modeling. It is recommended to start reading from the Introduction and Section 3, focusing on the calculation method of context vectors ci in equations (4) to (6), which is core to understanding how the attention mechanism works.

In the 1989 shareholder letter, Warren Buffett elaborated in depth on the core investment philosophy of Berkshire Hathaway. First, he distinguished between book value and intrinsic value. Although Berkshire's book value per share had grown at a compound annual rate of 23.8% over the previous 25 years, Buffett emphasized that "what matters is intrinsic value," defined as the discounted value of all future cash flows. He noted that as the company shifted from textile operations to a portfolio of higher-quality businesses, its intrinsic value was actually growing faster than its book value. This perspective reminds investors that accounting figures are merely a starting point; true value lies in a business's ability to generate cash flow.

Second, the letter detailed the mathematical logic behind long-term holding strategies, particularly the compounding effect of tax deferral. Buffett illustrated this with a hypothetical scenario: if stocks were bought and sold annually with capital gains taxes paid each time, even if the value doubled every time, the final return would be far lower than holding a single asset for 20 years and paying taxes only upon sale. He likened these unpaid taxes to an interest-free loan from the U.S. Treasury Department, which must only be repaid when the asset is sold. Therefore, reducing trading frequency is not just to avoid errors, but to leverage time value for more efficient compounding growth. Although he explicitly stated that choosing long-term holding is not solely based on tax calculations, but rather a desire to cherish relationships with excellent businesses, this mechanism objectively constitutes a significant competitive advantage.

Finally, Buffett offered profound insights into the unsustainability of high growth rates. Citing Carl Sagan's example of bacterial reproduction, he pointed out that in a finite world, high growth rates must self-destruct. As the base size increases, maintaining the same percentage of growth becomes extremely difficult. For Berkshire, achieving 15% growth on a net asset base of $4.9 billion is much harder than achieving 23.8% growth on a base of $22 million. This warns investors that as a company scales up, expected returns should be rationally adjusted downward, and one should not expect past high growth rates to continue indefinitely.

The "Berkshire Hathaway 1989 Shareholder Letter" was written by Warren Buffett and published in 1989 (the letter date is March 2, 1990). It is suitable for investors seeking to understand the underlying logic of value investing, focusing on long-term compounding mechanisms, and rationally viewing corporate growth cycles. It is recommended to start with the definition of intrinsic value at the beginning of the letter, then carefully read the tax comparison scenario analysis, and finally reflect on your own expectations for returns in light of the current market environment.