Summary

The core proposition of ultrasonic brain-computer interfaces is not to fully replace electrodes, but to trade slightly slower yet still closed-loop-capable temporal resolution for whole-brain coverage. 彭雷 estimates that Neuralink’s 1024 channels currently record roughly 1000–2000 neurons, covering only about 1.3‰ of the brain’s surface; an ultrasound probe of a similar size placed in the same location could cover roughly 25% of the brain’s volume. Electrical signals can be captured in under 10 microseconds, while the blood-flow signals ultrasound reads naturally lag neural activity by 0.5–1.5 seconds. But the approach opens a whole-brain path where “the whole is greater than the sum of its parts.”
Chronic pain is the ultrasonic indication closest to commercial realization, but regulatory approval remains the hard gate for device value. 彭雷 says a US partner uses a noninvasive phased array to stimulate the ACC, reducing pain immediately by 60%–70% for some patients after about 40 minutes, with effects lasting seven days. That could reduce dependence on painkillers, particularly for cancer pain, postherpetic neuralgia, peripheral neuropathy and phantom-limb pain. He expects approval in China and the US no earlier than late 2026 or 2027: “Without that certificate, your device is worth nothing.”
Stroke, AD, sleep and psychiatric disorders form a vast pipeline, while also exposing neuroscience’s evidentiary risk of knowing what happens without knowing why. Ultrasound may reduce neuronal death after ischemic stroke and open the blood-brain barrier to aid drug delivery; animal studies have also observed enhanced hippocampal metabolism and neuroregeneration. It may regulate circuits associated with insomnia, depression, ADHD, OCD, addiction and epilepsy. But there is still no consensus on whether Tau and Aβ are causes or consequences of AD. The real long-term value is not finding a few more targets where “poke it and something moves,” but gradually turning the brain from a black box into a “white box.”
Top-tier capital has pushed ultrasonic brain-computer interfaces from the fringe of research into an independent sector, but companies are still testing one indication at a time and at different levels of invasiveness. Players named on the show include Spire, backed by 陈天桥; Nudge, which raised a $100M Series A and has roughly 10 former Neuralink employees; Forest, supported by a donation from Eric Schmidt; Sanmai, backed by Reid Hoffman; and Merge Labs, co-founded by Sam Altman. The latter was reportedly in talks with OpenAI for a $250M financing round. 彭雷 summed up Nudge’s cross-industry appeal with a joke: “These web 3.0 people want to invest in web 4.0.”
The true long-dated option in ultrasound is to bypass the skull and combine the technology with genetic modification to create a scalable, whole-brain bidirectional interface. Because the skull causes person-specific scattering, 彭雷 envisions a “semi-invasive” procedure that leaves the dura intact, replaces part of the skull with resin, and combines ultrasound with gene therapy. He estimates this could still be 5–8 years away. The view is highly forward-looking and unvalidated, but the ceiling is clear: “Your entire brain could be read and written.”
Brain-computer interfaces, embodied intelligence and AI may eventually converge, but controlling robots and uploading knowledge or memories sit on entirely different maturity curves. Once electrodes decode a mouse, expanding to a robotic arm and eventually a bipedal robot is mainly a matter of adding degrees of freedom, and 彭雷 agrees with Musk that this could happen by 2028. But he rejects the idea that knowledge and memory uploads will be possible by then, saying local electrodes “will never be able to decode consciousness.” A language model running only on H200s will not develop consciousness, he argues; AI would first need a body, interaction with the physical world, and experiences of reward and punishment.
Chinese brain-computer and life-science companies benefit from lower clinical costs and denser resources, but they also need more patient capital, while geopolitics imposes a valuation discount. 彭雷’s baseline is that a US active implantable device in the relevant class takes an average of $1B and 10 years, while Chinese electrical brain-computer companies prepare on roughly RMB700M and 7 years. A single clinical case in China may cost only several hundred thousand yuan, versus roughly $1M–$1.5M for Neuralink in the US. Domestic electrical brain-computer companies average about $300M in valuation, compared with roughly $12B for Neuralink. His conclusion is that hard-tech companies in China and the US are already “being valued on separate scales,” but China’s talent, supply chain and patient resources leave room for long-term valuation recovery.

Deep dive

1. Local electrodes can treat disease, but cannot explain the whole brain

彭雷 starts with the brain’s 86B neurons to explain the traditional brain-computer interface: electrodes sit on the cortical surface or penetrate the cortex, recording neural firing to “read” and stimulating neurons to “write.” The approach has accumulated 30–40 years of research, and its clinical value for local decoding and modulation is established.
Parkinson’s is the clearest example of a local therapy: stimulate the STN, and a patient’s tremor can stop immediately. 彭雷 does not dispute that value, but distinguishes treating disease from answering ultimate questions about thought, consciousness, memory, emotion, and subjective and objective experience. The latter cannot be reconstructed by simply stitching together a handful of local signals.
After working on the electrical approach for 3–4 years, he founded Gestalt on a long-term hypothesis: “The brain must be something where the whole is greater than the sum of its parts.” Even if the cortex were packed with Neuralink electrodes, the result might still be an accumulation of local signals rather than the brain’s true operating mechanism.

2. fMRI sees the whole brain, but loses the timing that defines neural circuits

From X-rays, CT and PET-CT to MR, medical imaging has let humans noninvasively see gray matter, white matter, hemorrhages and tumors. But these are primarily structural signals. fMRI was the first to provide a whole-brain functional view, and over the past decade has produced maps of emotion, depression, autism and connections between brain regions.
The problem with fMRI is that it reads an indirect signal: rising blood oxygenation in a region. Even as magnetic fields have risen from 3T to 7T and spatial precision has improved, acquisition, reconstruction and the physiological gap between blood oxygenation and electrical activity can still create delays of several seconds or roughly 10 seconds.
Timing matters because brain regions have relationships of excitation, inhibition and strict sequence. 彭雷 compares the system to a city’s network of traffic lights: “Without a concept of time,” if adjacent signals were off by several minutes, traffic would collapse. Knowing only which regions lit up cannot reconstruct how a neural circuit operates.

3. A “transparent brain” ultimately requires approaching single-cell and 10-microsecond resolution at once

彭雷 defines the ultimate measurement target along two axes: locating all 86B neurons in space and recording when they fire in time. The program discusses targets ranging from the seconds level down to 100 milliseconds, with the long-term goal of reaching the tens-of-micrometers scale of individual neurons and the 10-microsecond scale.
A single neuron may fire more than 100 signals per second, with intervals between adjacent firings reaching the microsecond range. If one knew “what every cell did every 10 microseconds,” 彭雷 says, the brain would be close to transparent. That is a direction for the next 30 or 50 years, not a current product specification.
This framework also changes how technologies should be compared: electrodes excel at time resolution, fMRI at whole-brain spatial coverage, and ultrasound is trying to sit between the two while combining coverage, speed and bidirectional modulation into one platform.

4. Neuralink’s 1024 channels prove its temporal advantage while exposing its spatial ceiling

彭雷 says that since Neuralink implanted its first patient, Noland, in January 2024, the company had implanted 12 patients by the time of recording, all with 1024-channel devices. Each contact records an average of 1–2 neurons, but after accounting for insertion damage, drift and failures, he estimates that only about 1000–2000 neurons are actually readable.
Against 86B neurons, that is still negligible. The electrodes cover roughly 1.3‰ of the brain’s surface: “999‰ of the entire brain’s surface area has still not been recorded.” The larger physical constraint is that current electrodes generally penetrate only 3–5 millimeters, while the brain can be roughly 8 centimeters thick.
Its temporal performance is close to real time: once an electrode contacts a neuron, firing can be recorded within 10 microseconds. Musk has said the plan is to develop deep-brain electrodes and 25,000 pathways by 2028, as well as next-generation electrodes capable of reaching 6–7 centimeters, but those devices do not yet exist.
Chinese electrical brain-computer companies are mostly at 64–256 channels. 彭雷 considers maintaining an approximately 18-month doubling cycle excellent, because chips, communications, compute, volume and heat must all scale together like the staves of a barrel. He estimates China is roughly 3 years behind Neuralink, with the gap narrowing.

5. Ultrasound trades blood flow for coverage, potentially jumping from 1.3‰ to 25% at the same probe size

彭雷’s calculation is straightforward: if an ultrasound probe were shrunk to roughly the coin-sized footprint of a Neuralink implant and placed in the same location, Neuralink would cover about 1.3‰ of the brain’s surface, while ultrasound might image roughly 25% of the brain’s volume. This remains an engineering target, not a current product specification.
Ultrasound does not directly read electrical signals. It reads changes in blood flow through the microvessels and capillaries surrounding neurons. Neural activity requires blood supply, and the hemodynamic response can lead or lag across different neurons, creating an inherent physiological gap of roughly 0.5–1.5 seconds between blood flow and electrical activity.
泓君 asked whether that delay meant the device itself was slow. 彭雷 separates two issues: once blood flow changes occur, ultrasound can collect them in real time; what is slow is the mapping between the underlying physiological variables. Compared with fMRI’s several-second or even 10-second lag, ultrasound may still provide a whole-brain dynamic window centered roughly one second after the event.

6. Chronic pain brings noninvasive ultrasound closest to patients and revenue

Electrical brain-computer interfaces currently focus mainly on “reading”: decoding movement, speech or mouse control for patients with ALS and high-level spinal cord injuries to restore “Digital Autonomy.” Ultrasound’s earliest product opportunity is more likely to involve “writing”—using a phased array to focus acoustic energy on a specific brain region.
A phased array uses multiple elements to emit ultrasound at different timings and amplitudes, creating a focus inside the skull. Mechanical vibration or heating can excite or inhibit neurons. The principle is similar to phased-array radar, except that the target is brain tissue rather than a detected object.
The pain application targets the ACC: cancer pain, shingles-related pain, peripheral neuropathy and phantom-limb pain may ultimately converge on a central target. 彭雷 says early clinical work by a US partner showed that about 40 minutes of stimulation could reduce pain immediately by 60%–70%, with the effect lasting seven days.
The treatment requires no craniotomy. The patient lies in a hospital bed, secured in a head frame, and receives roughly 30–40 minutes of treatment. For people dependent on painkillers or whose drugs have lost effectiveness, it could mean “a week off pain pills.” But 彭雷 did not present the early result as an approved, universal therapy.

7. Safety parameters and regulatory approval—not engineering demos—determine when ultrasound becomes a business

Ultrasound has been used in B-mode imaging for decades and has advantages such as no radiation. High-energy ultrasound can also perform ultrasound ablation, aesthetic “ultrasound cannons” or tissue fragmentation. 彭雷 mentions a high-intensity ultrasound device used in Hong Kong for liver cancer that could liquefy tumors for the body to metabolize; that destructive power also shows why “hitting the brain” carries greater risk.
The research focus over the past decade has been finding the range of energy, focal points and parameters that can affect neurons without damaging other brain tissue. The FDA and NMPA are only beginning to accept ultrasound for brain modulation, and chronic pain has not yet received approval anywhere globally.
At the time of recording, 彭雷 judged that approval in China and the US might come as early as late 2026 or 2027. Registration must demonstrate both safety and efficacy: Class II devices usually take 3–4 years, while Class III devices can take 7–10 years. Medical products cannot follow the internet cadence of “one iteration every two weeks.”
He also notes that Neuralink’s 12 cases remain part of clinical research rather than an officially approved service. Ordinary patients are unlikely to have the option of hospital implantation before 2029 or later. Until then, participants are “research subjects,” not patients receiving standard treatment.

8. The disease pipeline is broad, but most opportunities remain in early validation

The acute or recovery phase after ischemic stroke is a second candidate: ultrasound may modulate damaged areas and reduce the mass death of neurons caused by interrupted blood flow.
The AD pathway is more varied. Ultrasound may accelerate the metabolism of Tau and Aβ deposits, increase microglial metabolic activity, or locally open the blood-brain barrier so drugs that normally cannot cross the vessel wall can enter brain tissue directly.
In animal studies, the team has also tried directly stimulating the hippocampus, observing enhanced metabolism and neuronal activity, and even possible “Neuroregeneration.” 彭雷 stresses that this has not yet been tried in humans. Some hippocampal neurons can regenerate, but whether that translates into an AD treatment remains unproven.
Sleep, severe depression, autism, ADHD, OCD, addiction and epilepsy are also potential applications. In early experiments, some people with chronic insomnia who were taking large amounts of sleeping pills fell asleep on the treatment bed. 泓君 went further, imagining high-quality sleep being reduced from 8 hours to 6; 彭雷 said this remained a large indication space. But targets vary by individual, and safety, generalizability and efficacy still need validation.

9. China started later, but may use patients and clinical resources to catch up on validation speed

彭雷 says some universities and research institutes in the US, Europe and Canada began experiments in parts of ultrasound brain modulation roughly 3–4 years ago, while China has only just entered the clinical research phase and Gestalt is still at the starting line. Because the sector itself is only about 10 years old, the China-US gap is not as wide as it is in mature device categories.
China’s first advantage is patient volume: a single indication may have a patient population 10 or 20 times larger than that of an individual European or American country. Its second is that nonhuman-primate experiments are relatively easier to conduct. 彭雷 therefore believes China could accumulate safety and efficacy data faster and might even secure a medical-device license first—but he says it is “entirely possible,” not an established outcome.

10. Silicon Valley capital has placed five types of ultrasound bets across different indications

Among the early players 彭雷 lists, Spire originated at the University of Utah and received investment from 陈天桥; Gestalt itself was also founded by 陈天桥 and 彭雷. The capital is betting not just on one device, but on whether ultrasound can become a new platform alongside the electrical approach.
Nudge has roughly 10 former Neuralink employees and completed a $100M Series A two months before recording, led by Thrive Capital. Its CEO Fred is a co-founder of Coinbase. The company chose noninvasive low-intensity ultrasound and PTSD as its entry point. 彭雷 joked that this was “web 3.0 people wanting to invest in web 4.0.”
Forest was founded by a Caltech postdoc and several professors. It follows the FRO, or Focused Research Organization, model, with former Google CEO Eric Schmidt as its main donor. It is more semi-invasive, placing a device beneath the skull and exploring arousal in patients with insomnia and disorders of consciousness.
彭雷 also mentioned Sanmai, backed by Reid Hoffman, and the newly formed Merge Labs. According to public information at the time, the latter was discussing a $250M financing with OpenAI; 彭雷 interpreted this as Sam choosing a technical route different from Elon Musk and Neuralink.

11. Merge Labs puts ultrasound and genetic modification on the same roadmap

Merge Labs has not disclosed a detailed plan, and 彭雷 says its website was inaccessible at the time. The two concepts visible so far are ultrasound and genetic modification. Gestalt is exploring a similar direction; in May, 彭雷 traveled to the US and discussed using genetic modification and protein design with David Baker to enhance related capabilities, but did not elaborate because “the frontier is sensitive.”
His basic assumption is that if the real goal is high-throughput connectivity between the human brain and AI—or even consciousness entering digital space, knowledge downloads or memory uploads—improving external devices alone may not be enough. “Our brains themselves will have to be modified.” Gene therapy is viewed as one possible way to improve the brain’s compatibility with a future interface.
彭雷 uses the red and blue pills in The Matrix as an analogy: the pills might be reinterpreted as gene therapies that alter the brain and connect it to the Matrix. He stresses that this is his own speculation and he does not know whether it is correct, but he likes the concept.

12. The skull is ultrasound’s biggest obstacle; a resin skull defines the semi-invasive ceiling

The “biggest obstacle” to reading and writing through the skull with ultrasound is bone. Differences in skull thickness and cavity structure from person to person create irregular scattering, reduce focusing precision and turn algorithmic processing into a complex problem.
Once part of the skull is removed, cerebrospinal fluid, the dura and brain tissue do not create major scattering or signal distortion for ultrasound transmission. 彭雷 therefore emphasizes that ultrasound can be semi-invasive while leaving the dura intact; Neuralink must open the dura and insert electrodes, creating more stringent safety and immune requirements. He compares replacing a piece of skull to rhinoplasty or chin reduction, while 泓君 points to the public’s intuitive fear of craniotomy.
The long-term concept is to replace portions of the skull on both sides with resin while preserving normal appearance, scalp and hair growth, allowing multiple probes to cover the whole brain. Combined with genetic modification, this could form a whole-brain read-write platform. 彭雷 estimates it is still 5–8 years away and says he considers the direction certain.

13. The brain’s 25 watts, tens of thousands of connections and plasticity expose the gap with GPU architecture

彭雷 compares H200 with the human brain: roughly 100B transistors and roughly 100B neurons are in the same order of magnitude, but each neuron can have 10⁴ connections. His power comparison is more striking: a GPU system consumes thousands or even tens of thousands of watts, while the brain uses about 25 watts.
The fundamental difference is not just process technology but structure. Mainstream computing uses a von Neumann separation of storage and compute; DeepSeek can run on Nvidia or Huawei hardware. The brain integrates storage and compute, so hardware and software cannot be separated. That is why biological computing is called “Wetware.”
Plasticity matters even more. 彭雷 says a single 2-hour conversation physically changes the synaptic connections, firing frequencies and connection strengths in the listener’s brain: “Your brain has already been permanently changed by me.” A manufactured AI chip cannot alter its internal wiring.

14. AI scientists see neuroscience and life science as the next sources of breakthroughs

彭雷 calls neuroscience and AI “two sides of the same coin.” Neurons inspired neural networks, and future breakthroughs in brain mechanisms could in turn reshape AI. He is particularly interested in whether brain-inspired chips, in-memory computing and other nontraditional architectures can reduce power consumption.
Jeffrey Hinton has a neuroscience background and raised a question 彭雷 considers important: modern AI converges through backpropagation and gradient descent, but biological brains do not use the same mechanism. If biology learns through another process, current AI theory has clearly not exhausted the nature of intelligence.
Demis’s work extends from AlphaGo and AlphaFold to AlphaGenome and the virtual-cell AIVC. 彭雷 explains that if AI could simulate how cells respond to perturbations, it might eliminate 80%–90% of unreliable compounds before they enter animal testing, then gradually expand to virtual organs and virtual humans.
He sees Isomorphic Labs as a commercial example of this convergence. The company has raised $650M, has four Nobel laureates among its advisers, and uses the slogan “Cure All Disease.” 彭雷 says many domestic discussions still frame AI around the internet, models and hardware, while some of America’s top AI scientists are redefining the problem through life science.

15. Language models have compressed knowledge, but still lack the physical world needed for consciousness

彭雷 believes the leap in large language models rests on the fact that most human knowledge is carried by language. Transformer, Attention, Scaling Law, training data, GPUs, reinforcement learning and synthetic data together drove the path from the Transformer paper to ChatGPT-5.
But language is a high-density, precisely describable symbol system. The physical knowledge behind opening a door, lifting a phone or walking has not been fully structured in world models. 李飞飞’s World Labs is cited as an example: next-generation AI must learn physics, chemistry and environmental feedback rather than merely traverse text.
On whether current AI models will develop consciousness, 彭雷 says that if AI remains at the stage of stacking H200s, it will not become conscious. It needs a body, interaction with the physical world, and reward and punishment before consciousness becomes possible.
The resulting convergence thesis is that AI, embodied intelligence and brain-computer interfaces will eventually merge. World models provide environmental understanding, robots provide bodies and action, and brain-computer interfaces may connect biological intelligence.

16. Speech can be decoded; semantics and memory remain distributed puzzles

Humans know that Wernicke’s and Broca’s areas are involved in speech and semantic processing, but 彭雷 admits: “We do not know why the brain produces language.” When someone expresses the same thought in Chinese or English, speech-related activity may change, while the concepts behind “桌子,” “chair” and “table” may share a semantic network.
Current electrical and ultrasonic brain-computer language decoding mostly remains at the level of speech or articulator movement. The sounds produced by the muscles of the mouth, pharynx, tongue and lips can be reconstructed, but that is not the same as reading abstract meaning. Semantics also involves the prefrontal cortex and distributed networks, again supporting whole-brain research rather than inserting devices into a few language areas.
Memory is likewise not a video stored in one location. The hippocampus is mainly involved in encoding, after which it replays the multimodal state of the experience. A week later, recalling a recording studio may reactivate the visual cortex for the teacup and surroundings, the auditory cortex for sound, and taste and touch regions as well. The memory is distributed across multiple brain regions.
彭雷 believes AI has not truly exploited this distributed replay mechanism and does not know how to do so. The hippocampus does not record an entire environment in one place like a camera; it participates in memory retrieval and encoding. That is also why there is still no clear engineering object for uploading knowledge or memories.

17. Medicine can work inside a black box, but scientific breakthroughs require illuminating the causal chain

The history of DBS for Parkinson’s shows that an effective therapy does not require a complete mechanism first. Doctors happened to find that removing a certain brain region stopped tremors, then reproduced the effect through destruction or electrical stimulation. DBS has been sold for roughly 20 years and may have been implanted in hundreds of thousands of people globally, yet the “why” remains incompletely answered.
彭雷 describes the current state as repeatedly “poking” a black box. If the same result appears more than 90 times out of 100, researchers establish a causal link and turn it into a drug or device. Ultrasound’s scientific value is that it could improve both whole-brain coverage and signal acquisition speed, giving researchers a chance to observe how stimulation affects broader networks rather than only local endpoints; its blood-flow signal, however, still differs physiologically from electrical activity by 0.5–1.5 seconds.
泓君 raised a key counterexample through AD: some patients have no Tau protein, and animal models may show AD-like symptoms without Tau. 彭雷 acknowledges that Tau, Aβ and even TBK, which appeared on the show, have not produced a unified target consensus. Whether deposits are causes or consequences remains unknown; sleep and psychiatric disorders have similar individual variation.
Whole-brain brain-computer interfaces are therefore closer to foundational life-science instruments than to a single therapy. 彭雷 compares them to gene sequencers: sequencing itself may not directly treat disease, but it first reveals how life is encoded and inherited, paving the way for precision medicine and gene therapy.

18. Medical devices must be financed on a 7–10-year timeline, while China’s cost advantage creates upside

彭雷’s minimum expectation for life-science entrepreneurs is 5 years and 7 years, not internet-style rapid iteration. His industry benchmark is that a Class III active implantable device in the US takes an average of $1B and 10 years, while Chinese electrical brain-computer companies should prepare for roughly RMB700M and 7 years to approval.
Clinical costs vary sharply by geography. A single case in China may cost several hundred thousand yuan, versus roughly $1M–$1.5M for Neuralink in the US. The expense comes from hospitals, physicians, surgery, post-operative care, rehabilitation, follow-up and insurance, not simply from manufacturing the implant.
Valuations have not been set according to that cost advantage. 彭雷 says domestic electrical brain-computer companies average roughly $300M, versus about $12B for Neuralink; Cambricon is worth roughly RMB600B, while Nvidia is worth about $3T. His conclusion is that Chinese and US hard-tech companies are now “being valued on separate scales,” with the early Alibaba-versus-Amazon comparison no longer serving as a common reference.
泓君 proposed the NewCo model for innovative drugs: since 2023 and 2024, use a US sales system or CEO to take on assets, conduct R&D and early clinical work in China, then export successful drugs to the US. 彭雷 believes biopharma and life science still retain room for license-in and license-out cooperation, unlike AI chips and quantum technology, which face tighter restrictions.

19. Robot control is visible engineering; consciousness upload remains an unexplained leap

Once an electrical brain-computer interface can control a mouse, it can theoretically control a robotic arm connected to a computer. More channels mean more decoded degrees of freedom, moving from a 2D cursor to a 3D or 6-DOF robotic arm and eventually a bipedal robot. 彭雷 agrees that Musk could achieve this by 2028; the main variables are speed and performance.
His stronger objection to Neuralink concerns knowledge and memory uploads by 2028: “With this kind of local electrode implantation, there will never be a way to decode consciousness.” Consciousness is the result of whole-brain, multi-region circuit regulation; it will not emerge simply because devices are installed in motor, visual, auditory and deep-brain areas.
Ultrasound currently shows only a higher ceiling for coverage; it has not explained the principle of uploading anything. 彭雷 is open to using enhanced interfaces in the future, but sees carbon-silicon integration in 2035–2045 as a possible choice: people might exist as robots retaining their memories, or retain only the brain while replacing the body’s organs.
His position is that of “an optimistic technological-progress humanist,” while acknowledging that death may no longer remain equal. People with more resources could gain longer lives and higher intelligence, deepening social stratification. Brain-computer interfaces are not the only cause; AI itself will create similar institutional and ethical pressures.

20. Six startups took 彭雷 from chasing internet waves to a 15-year bet on brain science

彭雷 describes himself as a “curiosity-driven serial entrepreneur.” His first startup was Bowei.com, an education-network content community launched while he was at the University of Science and Technology of China. It briefly ranked around 300 on Alexa, but he did not know how to monetize it and eventually sold it for tens of thousands of dollars around 2003 or 2004.
His second startup, Mingpian.com, was an SNS business that raised roughly $3M–$4M; he still owns the mingpian.com domain. The third was 24券, founded in 2010. During the thousand-group-buying war, it briefly ranked fourth or fifth and launched in the same week as Meituan. Group buying compressed a bubble cycle that normally lasted 7–8 years into 3 years, ultimately leaving Meituan as the sole winner.
Founded in 2012, Keruyun tried to complete the Offline To Online transition: rather than relying only on discounts to drive traffic offline, it digitized restaurant supply through an integrated hardware-and-software SaaS platform. The company received investment from Baidu, entered China’s NEEQ, and in 2019 chose to sell to Alibaba over Meituan, later becoming part of Ele.me, Koubei and Alibaba’s local-life ecosystem.
After leaving Alibaba in 2021, he was inspired by Elon Musk’s cross-disciplinary approach and first principles, and founded NeuroXess with a university classmate. He later set up Gestalt because the electrical route had become relatively established while he preferred “from 0 to 1.” Long-term brain-science exploration requires preserving short-term commercial revenue as well; this is his summary of Chinese entrepreneurs as “wildly imaginative yet grounded.”

21. Studying consciousness inward eventually leads back to humanity’s relationship with the universe

彭雷 uses Kant’s “starry heavens above” and “moral law within” to explain his two ultimate questions: humanity’s place in the universe, and why consciousness, emotion and value judgments exist. Carl Sagan’s idea that “the universe is a way for the universe to know itself through human consciousness” reconnects the two.
At this scale, brain science, Starship, quantum research and nuclear fusion are seeking different versions of the same answer: what is the universe’s purpose, and what does human existence mean to it? 彭雷 does not claim to have an answer. He only hopes neuroscience can contribute one small stretch of the path toward “the awakening of cosmic consciousness.”
The method of entrepreneurship ultimately returns to continuous learning. He moved from the internet and computer science into neuroscience and is now pursuing a neuroscience PhD at Fudan University. Four years ago he could not understand the papers; today he can speak with researchers and even Nobel laureates. He calls the process “my own large-model training journey, adding some new training data.”

Atlassian CEO, Mike Cannon-Brookes on Why Everything is Overvalued & Are We in an AI Bubble Michael Dell, Dell Technologies | David Senra