LesenForschungRadarAnlageframework
Anmelden / Registrieren
AlleTechnikProdukteInvestierenBörseSonstiges

Lesen

2026-10-30115 Beiträge

Aktualisieren

South Africa's Third-Largest Bank Absa Opens First Gulf Office to Tap Middle Eastern Capital

Thema · Absa海湾扩张Kurzfassung
2026-10-30 21:08 GMT+8

South Africa’s third-largest lender, Absa, has opened its first Gulf office, giving the bank a seat closer to Middle Eastern investors hunting for deals on the continent. According to Semafor, Absa CEO Kenny Fihla said the office would seek to benefit from advisory and credit-arrangement fees and help “connect capital from the region and from wider international markets with opportunities across Africa.”

Absa joins South African rivals Standard Bank, Nedbank, and Investec, which have already established a presence in Dubai. The move reflects a broader trend where Gulf Cooperation Council (GCC) countries have pumped more than $100 billion into Africa over the past decade. The United Arab Emirates is now Africa’s fourth-largest foreign investor behind China, the EU, and the US. As Gulf-backed ports, logistics, and energy projects expand along Africa’s eastern seaboard, continental banks are doubling down on East Africa and strengthening links with the Gulf.

Quellen:semafor.com

Forschung

Chinese memory-chip maker CXMT Q3 revenue jumps 613%, Micron and SanDisk shares tumble on competition fears

Thema · 长鑫长江跨界存储Wesentlich
2026-10-30 20:55 GMT+8

Chinese memory chip manufacturer ChangXin Memory Technologies (CXMT) reported its third-quarter financial results on October 30. The company saw its revenue surge by 613.2% year-over-year to 118.71 billion yuan ($16.7 billion), with net profit attributable to shareholders reaching 54.4 billion yuan ($8.17 billion), marking a significant turnaround from losses in the same period last year. These figures confirm earlier market expectations that demand for AI infrastructure has driven a sharp rise in memory chip prices.

CXMT attributed the improvement to higher memory-chip prices amid tight supply, alongside capacity expansion and a shift towards a more favourable product mix. As major rivals such as Samsung Electronics, SK Hynix, and Micron increasingly prioritize higher-value products like high-bandwidth memory (HBM), CXMT has benefited significantly from demand for conventional DRAM products.

The earnings announcement had an immediate impact on the global memory chip market. In premarket trading in the US, Micron Technology dropped over 6% and SanDisk fell nearly 9.5%. In South Korea, Samsung Electronics declined 3% while SK Hynix tumbled to close 7% lower. Investors are reassessing the durability of elevated memory margins against the backdrop of intensifying competition from China and potential oversupply risks.

Additionally, CXMT disclosed progress on a new DRAM cell architecture known as 4F², designed to increase storage density, with plans to introduce a server-memory product using this technology by the end of 2026. Although large-scale commercial production has not yet been demonstrated, CXMT's rapid financial growth and manufacturing expansion have heightened concerns about the changing competitive landscape in conventional DRAM.

Quellen:invezz.com

Forschung

Micron Gives Up to Five Years' Salary in Bonuses to Taiwan Production Staff Amid AI Memory Boom

Thema · 美光AI存储Wesentlich
2026-10-30 19:44 GMT+8

Amid the AI-driven memory chip boom, US-based semiconductor manufacturer Micron Technology is distributing massive bonuses to its production-line employees in Taiwan. According to Digitimes, as reported by Tom's Hardware, approximately 700 eligible Micron employees in Taiwan will receive a lump-sum payout equivalent to five years' worth of salary. This accounts for about 17% of the more than 4,300 production-line staff working at Micron Taiwan. The remaining employees will also receive three bonuses in October.

The lump-sum payout consists of three incentives: an annual performance bonus estimated between 7.5 and 12.5 months of salary for most workers, a mass-production bonus amounting to 20% of their annual salary, and a one-time $31,000 (NT$1,000,000) appreciation bonus, which will be prorated for employees who haven't reached a full year of service. Additionally, Micron workers globally expect a new quarterly profit bonus starting in January 2027, paid 50% in cash and 50% in stock, with the first payment being 100% in cash.

This move reflects the extreme profitability of the current AI memory market. Micron is following in the footsteps of its South Korean rivals, Samsung Electronics and SK Hynix, which have recently paid out massive bonuses to their workers of up to $400,000. As AI hyperscalers drive unprecedented demand for advanced storage products like High Bandwidth Memory (HBM), the big three memory chip manufacturers are raking in record orders and profits, leading production-line workers to demand a slice of the pie. For instance, Samsung workers voted to accept a $340,000 average bonus in May 2026, while SK Hynix workers are set to receive $477,000 in bonuses this year, potentially growing to almost $900,000 in 2027.

Despite these seemingly massive payouts, labor disputes persist. The workers' union at Micron Taoyuan is still in the midst of a dispute with management regarding pay. Some officials have already launched a one-day strike, calling for clearer profit-sharing arrangements and terms for bonuses, with nearly 70% of members at the Taoyuan site turning down management's latest offer. The two sides are tentatively set to meet again on November 6. On the other hand, the union at Micron Taichung has already agreed to a 10% employer contribution towards worker pensions and a tripling of their day-shift allowance.

For investors, these hefty bonuses reveal the current state of profit distribution in the upstream AI hardware supply chain. While such incentives help chipmakers stay competitive in the war for talent, their heavy reliance on performance and profit bonuses means that if AI demand falls or the industry cycle returns to normal, these incentive costs could precipitously drop, directly impacting the operating expenses and margin structures of these companies.

Quellen:Tom's Hardware

Forschung

Epoch AI Estimate: Global AI Chip Quarterly Compute Capacity Additions Double Every Nine Months

Thema · Epoch AIWesentlich
2026-10-29

Research organization Epoch AI released data insights on October 29, 2026, estimating the quarterly growth rhythm of global AI chip compute capacity additions. Its methodology involved using quarterly sales estimate data from six major chip designers—Nvidia, Google, AMD, Amazon, Huawei, and Cambricon—to perform a log-linear regression fit on ten quarters of data from Q1 2024 to Q2 2026, measuring capacity in H100-equivalent units (H100e). The results show that quarterly compute capacity additions grew from 800,000 H100e in Q1 2024 to 6.3 million H100e in Q2 2026. The fitted trend indicates an annual growth of 2.5 times (90% confidence interval: 2.3× to 2.7×), corresponding to a doubling period of 9.1 months (90% confidence interval: 8.4 to 10.0 months).

Epoch AI deliberately chose quarterly additions rather than cumulative stock as the subject of analysis, a methodological choice worth explaining. When additions grow at a stable rate, the growth rate of cumulative stock converges to that same rate; however, if sales surge and then slow, the two diverge. 2023 was precisely such a case: sales of Nvidia GPUs and Google TPUs rose sharply within the year, with H100e sales in Q1 2024 being approximately six times higher than in Q1 2023. Because cumulative stock remained small relative to quarterly sales at that time, even with stable growth in additions, cumulative stock experienced large proportional jumps each quarter due to base effects. By 2026, the two indicators had converged: cumulative capacity grew 2.6 times between Q2 2025 and Q2 2026, close to the 2.5× fitted trend of quarterly additions, down from 3.5 times the previous year. Epoch AI therefore considers that fitting the trend to cumulative stock for 2024–2026 would overstate the current growth pace.

Regarding whether growth is slowing, the signals from the data are ambiguous. Year-over-year growth in quarterly additions fell from 2.8 times in Q2 2025 to 2.2 times in Q2 2026; regressing quarterly growth rates against time shows a downward trend from approximately 32% to 19%, but this is not statistically significant (p = 0.26). Epoch AI points out that quarter-to-quarter fluctuations are inherently large, partly due to chip generation transitions—for example, the shipment transition of Nvidia’s Blackwell chips from late 2024 to early 2025.

This estimate has clear boundaries. The data tracks the point of chip sale rather than the point of going online: chip sellers like Nvidia typically recognize revenue for AI chips when they deliver them to server assembly partners, meaning these chips may take several more months to actually operate in data centers. Additionally, the analysis does not include smaller designers such as Intel, Meta, Microsoft, Cerebras, and Groq, nor GPUs intended for non-data-center uses. Epoch AI believes these volumes are insufficient to alter the trend. The confidence intervals only reflect how well the trend fits the median sales estimates, excluding uncertainty in the underlying sales estimates themselves. The ten-quarter window is also relatively short, and chip sales tend to advance in waves as new generations ramp up.

Quellen:Epoch AI

Forschung

Epoch AI says open-weight model GLM 5.3 autonomously exploits Firefox patch gap for ~$1,000

Wesentlich
2026-10-30 20:22 GMT+8

According to The Decoder, researchers at Epoch AI found that the widely available open-weight AI model GLM 5.3 can autonomously exploit the 'patch gap'—the window where security fixes are visible in code but have not yet reached users—to attack Mozilla Firefox.

In testing, the model built full exploit chains against six Firefox versions released between September 17 and October 26, 2026. It identified disguised security fixes, selected vulnerabilities, and chained them with a previously unpublished memory safety bug to bypass sandbox protections and read local files. This process required zero human input, taking as little as 12 hours and no more than 48 hours.

Each successful attack cost approximately $1,000 in API fees. Researchers noted that while framed as a security exercise, the actions were indistinguishable from those of an attacker targeting real users, and neither Z.ai nor other API providers triggered any safety guardrails or refusals.

This finding suggests that advanced cyber capabilities are diffusing through open-weight models. Although the team did not produce an equivalent exploit against Google Chrome using the same setup, they cautioned this does not establish immunity, warning that similar risks exist across the software industry, particularly for infrequently updated systems like router firmware and mobile operating systems.

Quellen:The Decoder

Forschung

Pseudonymous Software Engineer Morluto Creates AI Reverse-Engineering Tool REA, Enabling Ordinary Users to Deconstruct Device Proprietary Software

Thema · REA AI逆向工程工具Wesentlich
2026-10-30 19:30 GMT+8

For decades, software manufacturers have enjoyed a sense of security derived from tedium: dismantling compiled programs was so exhausting that almost no one bothered. But this packaging also kept users out of the inner workings of their own devices. If you buy a new car, you can repaint it or swap the wheels, but you are not allowed to modify the control code for the dashboard display and other electronics. In an exclusive interview published by Scientific American on October 30, 2026, a software engineer using the pseudonym Morluto viewed these restrictions as a challenge. His tool, REA (Reverse Engineer Anything), allows AI agents to autonomously orchestrate decompilation programs, restoring binary code into human-readable source code, and then reassembling it according to the user's wishes.

The technical principles behind REA are not mysterious. Before a program runs, its code must be compiled into binary—a long string of 0s and 1s that instructs the computer on how to operate. Reverse engineering involves reversing this translation, converting binary back into human-readable code during the decompilation process. Previously, reverse engineers manually performed this work using tools such as Ghidra, developed by the U.S. National Security Agency. REA provides the pipelines that allow AI agents to call upon Ghidra and other decompilation programs independently. "Reverse engineering is somewhat like pattern matching, and AI possesses extremely strong reasoning and pattern-matching capabilities," Morluto said.

The lowering of barriers is the core change highlighted in this article. Zion Basque, an assistant professor at the University of Georgia who studies reverse engineering, noted that before 2026, decompilation was the domain of a select few experts; now, nearly anyone can participate. His research team developed Kuna, one of the decompilation tools that REA can equip AI agents with. James Grimmelmann, a law professor at Cornell Tech at Cornell University, stated that completing such tasks manually would require experts months of work, if they could manage it at all; the current limitation is simply how many tokens you can afford. "If you have the tokens, you can set it loose on almost any piece of software."

The diffusion of this capability is shaking up legal frameworks. Grimmelmann believes that copyright protects the program's code but not the algorithms and ideas implemented by the code. "This opens up a world of mods, remixes, and interoperable new clients, significantly shortening the lead time people gain from deploying software." He warned, "This will pose a crisis for trade secret and copyright law."

For ordinary users, this capability means gaining control over the devices they have purchased. Joey Castillo, an open-source developer and technologist in residence at Cornell Tech at Cornell University, previously attempted to modify a smart ring he bought in Shenzhen, China. He discovered that AI models could not only reverse-engineer the ring but also teach him the entire process simultaneously. "I learned a lot just by watching Claude work," he said. Castillo attributed this desire to tinker with a need for control over items he had bought: "I bought this object, and it is being gatekept from me. Now I want to make it do what I want."

Quellen:Scientific American

Forschung

Pepsi Writes Down $300M Earnout Fair Value to $12 Million: Why Earnout Clauses Systematically Fail to Pay Out

Thema · 并购对赌估值争议Wesentlich
2026-10-30 19:29 GMT+8

In May 2025, PepsiCo acquired the prebiotic soda brand Poppi for $1.95 billion in cash plus a $300 million earnout clause. The earnout stipulates that if Poppi achieves specific performance milestones by the third quarter of 2027, Pepsi must pay the $300 million; otherwise, no payment is made.

Pepsi’s latest financial report shows that the fair value of this $300 million earnout has been written down from a peak of $278 million to just $12 million. The author of OnlyCFO points out that this means Pepsi believes the likelihood of meeting the target is extremely low—$12 million represents only 4% of the $300 million target, and the clause structure is all-or-nothing.

It is virtually certain that the sellers of Poppi will not receive any money from this earnout.

The original intent of an earnout clause is to bridge valuation gaps between buyers and sellers. The buyer says: 'If you believe the company is worth more, then prove it with performance.' But once the transaction is complete, the seller loses control over the outcome. The buyer decides on hiring plans, marketing budgets, and sales incentives—these are precisely the key variables that determine whether performance targets can be met. Speaking from the perspective of a buyer, the author admits: there is no malice, but priorities change. Budgets for acquired products are often reduced, and sales teams may lack the motivation to push new products, making targets suddenly unattainable.

Another common trap lies in metric definitions. How is Annual Recurring Revenue (ARR) calculated? Which costs are included in gross margin? Does this revenue count towards churn? Unless the agreement is written with extreme clarity, the buyer almost always wins in disputes over definitions. The author specifically notes that confusion over the definition of ARR is a major source of earnout disputes in the software industry.

Industry data confirms this structural imbalance. According to SRS Acquiom, approximately 25% of recent deals include earnout clauses, but 41% of these earnouts ultimately result in zero payment, with only 21% of the maximum earnout amount actually paid out. Sellers contested the earnout in 28% of these transactions, and after contestation, only 17% saw an increase—the buyer usually prevails.

From an accounting perspective, changes in the fair value of earnout clauses directly impact the income statement. Pepsi increased the fair value of the earnout to $278 million in the fourth quarter of 2025, resulting in a $98 million expense; subsequently, it was reduced each quarter: to $162 million in Q1 2026 ($116 million gain), to $117 million in Q2 2026 ($45 million gain), and to $12 million in Q3 2026 ($105 million gain). If no payment is ultimately required, Pepsi will cumulatively recognize a $180 million gain. In other words, the failure of the earnout to pay out actually makes Pepsi’s income statement look better.

The author’s conclusion is blunt: do not count on the earnout. Treat the upfront cash as the actual receipt, and view any earnout as a bonus. If you must accept an earnout structure, you should demand covenants regarding product budgets, clear metric definitions, monthly or quarterly tracking, a shortened assessment period of 1-2 years, inquiries into the buyer’s historical earnout payment records, and strive to exchange earnout potential for more upfront cash—as, on average, only $0.21 of every $1 in earnout potential is actually paid out.

Quellen:onlycfo.io

Forschung

GPU Cloud Provider GMI Cloud Showcases Taoyuan AI Factory: Power Replaces Space as Compute Expansion Bottleneck

Thema · GMI CloudWesentlich
2026-10-30 19:45 GMT+8

GMI Cloud, a provider specializing in GPU cloud computing, showcased the first phase of its AI Factory in Taoyuan, Taiwan, on October 30, 2026. The facility was not newly built but converted from existing Internet Data Center (IDC) infrastructure owned by Taiwan Mobile. The tour revealed that despite significant unused floor space, the core constraint on expanding compute capacity has shifted from physical area to power availability.

According to iThome's report, GMI Cloud's deployment engineers stated that the central question for AI-era data centers has changed from "how many more racks can we fit" to "how much power can support GPU computation." The site has a maximum load capacity of approximately 25 MW, but only about 16 MW is available for IT loads, with the remainder dedicated to cooling systems, air conditioning, and other auxiliary equipment. This advantage stems from previous cautious power management by Taiwan Power Company for large data centers north of Taoyuan due to grid constraints, making existing facilities with substantial power capacity more viable for AI deployment.

The first phase features Nvidia DGX SuperPOD GB300 systems, comprising three SuperPODs and roughly 48 AI racks. A second phase, planned for December, will add approximately 3.5 SuperPODs. GMI Cloud noted that future expansion cannot simply involve adding more racks; it requires solving power efficiency issues. If the IT load approaches the 16 MW limit, additional GPU equipment cannot be installed regardless of available space.

To ensure stability for high-density GPU operations, the AI Factory employs a multi-layered power backup architecture. Lithium batteries provide an immediate buffer for about six minutes during grid interruptions, followed by UPS systems maintaining power until diesel generators start. The N+1 designed diesel generators have fuel tanks capable of supporting over 24 hours of full-load operation. Additionally, transformers, batteries, and UPS units are distributed across the site to mitigate single-point failure risks.

Regarding thermal management, the facility uses a liquid cooling architecture that removes approximately 87% of heat, with air cooling handling the remaining 13%. Since the original IDC chilled water system operated at lower temperatures (16–17°C) than required for AI GPU liquid cooling (35–38°C), Coolant Distribution Units (CDUs) are used to integrate the systems via heat exchange. GMI Cloud reported a current Power Usage Effectiveness (PUE) of about 1.35, aiming to reduce it to 1.3 through optimized cooling controls. This case illustrates that competition in AI infrastructure is shifting from land and rack scale to effective compute output per unit of power.

Quellen:iThome

Forschung

Samsung Division Heads Mark 57th Anniversary: Semiconductor Chief Warns Against Complacency, Device Chief Prioritizes Cost Competitiveness

Thema · 三星AI零部件成本压力Wesentlich
2026-10-30 07:24 GMT+8

Heads of Samsung Electronics' semiconductor (DS) and consumer devices (DX) divisions issued distinct messages marking the company's 57th anniversary, highlighting the divergent impacts of the artificial intelligence boom on their respective businesses. Jeon Young-hyun, head of the DS division, warned against complacency despite historic performance, emphasizing that technological competitiveness has not yet fully recovered. No Tae-mun, head of the DX division, directly addressed the operational difficulties caused by rising component prices driven by AI demand, making cost competitiveness a primary objective.

Jeon noted that while AI market growth and expanding semiconductor demand have led to record-breaking results, "now is the time to calmly reflect on our reality and competitiveness and prepare for the future, rather than celebrating achievements." He admitted, "Although recent performance has improved, it is difficult to say that our technological competitiveness has sufficiently recovered." He stressed that "super-gap" leadership must be proven through customer-recognized technology and quality, not just slogans. Challenges cited include preparing for next-generation computing architectures, improving yields, and building a business system resilient to market fluctuations. With memory price surges strengthening supplier bargaining power, Jeon explicitly cautioned against leveraging this position to make unreasonable demands on customers, aiming to preserve long-term trust.

In contrast, No acknowledged the severe operating environment facing the DX division. He described the industrial shift triggered by AI as faster and more intense than the previous analog-to-digital transition, noting specifically that the surge in semiconductor demand has inflated prices, squeezing margins for end devices. To counter this, he called for enhanced proactive supply chain sensing and management capabilities to secure differentiated cost competitiveness. No identified product completeness and quality as fundamental strengths and pointed to new success stories in foldable phones, Micro RGB TVs, smart glasses, and AI companion devices. Additionally, he outlined plans to redesign business processes using AI to transform into an "AI-native company," reinvesting resources gained from efficiency into high-value-added tasks and new growth areas.

Quellen:미래를 보는 창 - 전자신문

Forschung

Nonprofit Prosper Global Says Agentic AI System VERA Cut Crisis Research Time in Half in Sudan

Thema · VERAKurzfassung
2026-10-30 20:02 GMT+8

Aaron Eubank, Director of Data Science at Prosper Global (formerly Mercy Corps), stated at the EVOLVE26 conference that the organization's multi-agent research engine, VERA, has shown potential to accelerate crisis response research in conflict zones. Built in partnership with big data company Cloudera, the system is designed to provide analysts with a faster starting point rather than automating the entire research process, ensuring researchers retain agency.

According to Eubank, VERA leverages Cloudera AI Studios, Amazon Web Services infrastructure, and models from Anthropic to pull information from sources such as ReliefWeb, the ACLED conflict tracker, and local news. In Sudan, the time required for desk research on an agricultural report was reduced from five or six days to two, saving approximately $1,500. In Colombia, an employee with no security background produced an election security report in just a few hours.

Launched in June 2026, VERA is also active in the Democratic Republic of Congo and Ukraine. Eubank noted that risks and contexts vary by location, necessitating a measured and slow approach. These specific metrics are self-reported by Prosper Global and have not been independently verified.

Quellen:SiliconANGLE

Forschung

South Korean Police Investigate 80 Tesla Drivers for Illegal FSD Unlock, Seize 21 Hardware Modules

Kurzfassung
2026-10-30 18:20 GMT+8

According to Tom's Hardware, the Seoul Metropolitan Police Agency has identified 80 Tesla drivers suspected of using unauthorized hardware to force-enable Full Self-Driving (FSD) on vehicles not cleared for the feature. The police reported confiscating 21 device modules used to switch on the feature. These unlock modules connect to the car’s CAN bus (the internal network linking electronic control units) to bypass the regional software lock that keeps FSD disabled. This is technically feasible because Tesla builds the same computer into every car, regardless of local feature approval.

Suspects imported ready-made units from Europe and China or built their own. One IT worker constructed a device using information found on GitHub, suggesting a low barrier to entry for such modifications. The group also shared tips online on how to conceal activation logs to avoid detection by Tesla. Under South Korea's Motor Vehicle Management Act, unauthorized software changes that could affect vehicle safety are illegal, carrying penalties of up to two years in prison or a fine of 20 million won (approximately $15,000).

The crackdown stems from a certification gap in Tesla's Korean market rollout. Tesla began offering FSD (Supervised) in South Korea in late 2025, but only on US-built Model S, Model X, and Cybertruck units, which skip certain local safety certifications under the Korea-US free trade agreement. Model 3 and Model Y vehicles imported from Giga Shanghai have yet to pass these requirements. South Korea's transport ministry first flagged suspicious activations to police in April, with officials confirming safety risks including potential failure to recognize driving conditions correctly.

Quellen:Tom's Hardware

Forschung

Foxconn Tech Day Reveals: NVIDIA Vera Rubin Compute Tray Assembly Cut to 20 Minutes, Some Product Development Takes Just 3 Weeks

Thema · 鸿海AI制造效率Wesentlich
2026-10-30 19:06 GMT+8

Taiwan-based electronics manufacturing giant Hon Hai Technology Group (Foxconn) held its annual Tech Day on October 30, 2026, where Chairman Young Liu and Rotating CEO Jimmy Chao disclosed several manufacturing efficiency metrics specifically targeting artificial intelligence infrastructure. The most notable figure revealed was that the assembly time for Compute Trays used in the NVIDIA Vera Rubin platform has been reduced from over two hours to just 20 minutes.

This efficiency gain is attributed to a "Cableless" solution co-developed with clients and fully automated assembly technologies. Young Liu noted that while outsiders often view Foxconn merely as a system assembler or "box mover," this year's event focused on showcasing its precision machining capabilities and vertical integration. In terms of precision machining, Foxconn achieved sub-micron accuracy using high-speed spindles and precision tools, applying this technology to critical components like optical communication chip connectors for next-generation AI servers.

Beyond assembly speed, Foxconn emphasized accelerated New Product Introduction (NPI) cycles. Data showed that in 2025, Foxconn executed 941 NPI projects, averaging about 2.5 per day. Jimmy Chao explained that by involving engineers from component, mold, and functional module domains early in the design phase, and utilizing cross-timezone relay collaboration between Taiwan and US teams, some product development cycles have been compressed from the industry-standard three months to just three weeks.

Regarding AI infrastructure layout, Foxconn stated its manufacturing scope covers Levels 1 through 12, forming a production network spanning Asia and North America, from component manufacturing in Taiwan and Vietnam to system integration in the US and Mexico. Additionally, Foxconn exhibited Spectrum-X Co-Packaged Optics (CPO) switches manufactured for NVIDIA, claiming that micron-level alignment technologies have progressed from laboratory stages to mass-producible processes, indicating readiness to address the thermal and connectivity challenges posed by increasing AI compute density.

Quellen:TechOrange 科技報橘

Forschung
Weiter →
Anmelden / Registrieren
LesenForschungRadarAnlageframework

However, some barriers exist for important reasons. While AI excels at breaking into security systems, it is also very adept at disrupting digital rights management and protected firmware within devices. Castillo and his open-source peers worry that people will meddle with devices without fully understanding them. "All of this will be very fun until someone vibe-codes firmware and causes a battery to catch fire mid-flight," Castillo said. He doubts that AI models possess the foresight that reverse-engineering experts bring when deciding which modifications to make, because AI "can only say 'yes, and,' not 'no, but.'"

Platform-level restrictions have already emerged. Since Morluto's initial interview with Scientific American, social media users have reported that OpenAI and Anthropic banned or blocked their use of REA. An OpenAI researcher responded on X, stating that the block on reverse engineering was "mostly accidental." However, Castillo already has alternatives: the open-weight model GLM-5.3 from the Chinese company Z.ai. "Once downloaded, no one can take it away," he said. "If Anthropic and OpenAI decide to ban my account, I will launch GLM-5.3 to reverse-engineer that smart ring, because I think it’s really cool to make this thing do what I want."

Morluto is not concerned about these obstacles. In his view, the tech industry is full of walled gardens, and AI has made "their walls not as robust as they thought." He repeatedly mentioned tech founders who shaped his work—Steve Jobs of Apple, Jensen Huang of Nvidia, and Paul Graham of Y Combinator. "Many creators in the past were hackers, tinkers, and experimenters growing up," he said. He believes REA can turn users into that kind of creator. "You can not only perform magic but also create beautiful things." Regarding the core of REA, Morluto said, "It is a path that guides you to question what you truly want."