Kimi released K3.1 Pro on October 24, cutting the API input price from 20 yuan to 13 yuan per million tokens and the output price from 100 yuan to 68 yuan, but users questioned the cost-effectiveness due to rapid token consumption; on October 26, the company apologized, compensated quotas, and adjusted the model configuration. The author used a Pro subscription account to test five creative tasks, which consumed a total of about 25.48% of the monthly quota, with "The Peach Blossom Spring" at 6.91% the highest and the mechanical keyboard at 1.46% the lowest; consumption mainly came from Plugin entries and multiple rounds of revisions. The article also mentions the controversy over overseas bloggers being asked to hide paid partnership labels in promotions, and that former miHoYo international president Jin Wenyi joined Moonshot AI to oversee overseas expansion.
Kimi released K3.1 Pro on October 24, cutting the API input price from 20 yuan to 13 yuan per million tokens and the output price from 100 yuan to 68 yuan, but users questioned the cost-effectiveness due to rapid token consumption; on October 26, the company apologized, compensated quotas, and adjusted the model configuration. The author used a Pro subscription account to test five creative tasks, which consumed a total of about 25.48% of the monthly quota, with "The Peach Blossom Spring" at 6.9
Karan Singhal, head of OpenAI's health division, said in an interview that more than 300 million people ask ChatGPT health-related questions every week, and that OpenAI chooses to incorporate medical data at each training stage of its general models rather than training a separate medical-specific model, to avoid the specialized model quickly falling behind frontier models. He mentioned that OpenAI is exploring a proactive personal health assistant covering before, during, and after consultations, and said users' health information will not be used for model training or ad targeting. OpenAI is also working with Penda Health to experiment with a clinician assistance system, and with medical information systems such as Epic to explore using ChatGPT directly within healthcare workers' existing work environments.
Karan Singhal, head of OpenAI's health division, said in an interview that more than 300 million people ask ChatGPT health-related questions every week, and that OpenAI chooses to incorporate medical data at each training stage of its general models rather than training a separate medical-specific model, to avoid the specialized model quickly falling behind frontier models. He mentioned that OpenAI is exploring a proactive personal health assistant covering before, during, and after consultation
At Zhiyuan's fourth Innovation Day, co-founder Peng Zhihui (Zhihui Jun) responded to the question raised by GPT-6 Astra's ability to directly control robots, arguing that regardless of whether foundation models can handle embodied capabilities across the board, body manufacturers that can scale, keep costs low, and achieve high consistency will be more advantaged. He cited data showing Zhiyuan shipped more than 5,000 units in 2025 and 9,700 units in the first half of 2026, and said that networked post-training with 100 real machines can raise task success rates from about 50% to 98%. Zhiyuan also proposed using Worken to measure a robot's effective intelligent work hours, and pointed out that the human-to-machine ratio in operations and maintenance is a key bottleneck for scaling.
At Zhiyuan's fourth Innovation Day, co-founder Peng Zhihui (Zhihui Jun) responded to the question raised by GPT-6 Astra's ability to directly control robots, arguing that regardless of whether foundation models can handle embodied capabilities across the board, body manufacturers that can scale, keep costs low, and achieve high consistency will be more advantaged. He cited data showing Zhiyuan shipped more than 5,000 units in 2025 and 9,700 units in the first half of 2026, and said that networke
Google's third-quarter search revenue grew about 12%, below analysts' expectation of 15%, and some investors believe the problem of AI search queries growing fast but having low monetization capacity is beginning to show. Executive Philipp Schindler said on the earnings call that the strong growth of AI Mode queries affected revenue due to the lower level of monetization. In the same period, cloud business revenue grew 103% year over year to $30.7 billion, and the Gemini app reached 1 billion users.
Google's third-quarter search revenue grew about 12%, below analysts' expectation of 15%, and some investors believe the problem of AI search queries growing fast but having low monetization capacity is beginning to show. Executive Philipp Schindler said on the earnings call that the strong growth of AI Mode queries affected revenue due to the lower level of monetization. In the same period, cloud business revenue grew 103% year over year to $30.7 billion, and the Gemini app reached 1 billion us
Madhulika Srikumar and Claire Leibowicz of Partnership on AI write that current agent frameworks cannot reliably generate the key observability records needed for safe deployment. Together with colleagues from academia, civil society, and industry, the authors studied the four most widely used agent frameworks and found that signals such as persistent Agent IDs and human approval records are missing, making it difficult to determine which agent did what and whether it exceeded its authority. The article recommends using OpenTelemetry to fill in the records of "what happened," and calls on framework developers, enterprises adopting agents, and regulators to each bear their respective responsibilities.
Madhulika Srikumar and Claire Leibowicz of Partnership on AI write that current agent frameworks cannot reliably generate the key observability records needed for safe deployment. Together with colleagues from academia, civil society, and industry, the authors studied the four most widely used agent frameworks and found that signals such as persistent Agent IDs and human approval records are missing, making it difficult to determine which agent did what and whether it exceeded its authority. The
Nathan Lambert reviews the newly released Intelligence Index v5 from Artificial Analysis, which weights knowledge work, coding, science, and general each at 25%, and presents new model rankings. He notes that 8 closed-source models in the rankings lead the best open-source model, Kimi K3.1 Pro, and argues that the new index leans more toward measuring where frontier models are heading, potentially underestimating the role of open-source models in the economy. The article also discusses the pros and cons of shifting to private benchmarks, and how the business model of benchmark organizations charging model vendors to verify results affects credibility.
Nathan Lambert reviews the newly released Intelligence Index v5 from Artificial Analysis, which weights knowledge work, coding, science, and general each at 25%, and presents new model rankings. He notes that 8 closed-source models in the rankings lead the best open-source model, Kimi K3.1 Pro, and argues that the new index leans more toward measuring where frontier models are heading, potentially underestimating the role of open-source models in the economy. The article also discusses the pros
Hugging Face released four talks, a panel discussion, and a Q&A, discussing practical approaches to training Agents in industry, with the core conclusion being that environment, evaluation, and reward design are harder than the model itself. Miles used 64×B300 fully asynchronous training on Terminal Bench 2, and after about 100 steps the reward rose from about 0.44 to 0.56; OpenEnv Arena, after filtering down layer by layer from about 90,000 findable RL environments, had only about 800 relatively clean and usable ones left. The talks also mentioned cheating problems such as Agents finding reference answers in the sandbox and modifying the scoring logic to fake a pass, as well as about 1% of cheating problems contaminating an entire round of training.
Hugging Face released four talks, a panel discussion, and a Q&A, discussing practical approaches to training Agents in industry, with the core conclusion being that environment, evaluation, and reward design are harder than the model itself. Miles used 64×B300 fully asynchronous training on Terminal Bench 2, and after about 100 steps the reward rose from about 0.44 to 0.56; OpenEnv Arena, after filtering down layer by layer from about 90,000 findable RL environments, had only about 800 relativel
It has been 130 days since WeChat's native AI assistant "Xiaowei" launched. Its entry points have expanded from the 16 officially disclosed to 28 measured by third parties, covering chat, search, official accounts, Channels, and PC. Through the developer integration guide released on June 8, WeChat requires developers to break mini programs into atomic interfaces and atomic components and write SKILL documentation; mini programs that have not integrated will be unable to be recommended and invoked by WeChat AI. The article also mentions that the development model must pass platform evaluation (80 points or above in the three items of service delivery, interaction experience, and performance quality, and 60 points or above in scenario coverage) to be eligible for invocation, while the invocation ranking rules have not yet been made public.
It has been 130 days since WeChat's native AI assistant "Xiaowei" launched. Its entry points have expanded from the 16 officially disclosed to 28 measured by third parties, covering chat, search, official accounts, Channels, and PC. Through the developer integration guide released on June 8, WeChat requires developers to break mini programs into atomic interfaces and atomic components and write SKILL documentation; mini programs that have not integrated will be unable to be recommended and invok
In a Latent Space podcast interview, Anthropic engineering lead Katelyn Lesse and product lead Angela Jiang discussed the Opus 5.5 release, covering adjustments to the model's writing style, pricing based on intelligence per dollar, and API changes such as requiring thinking blocks to be retained across turns to prevent distillation. The interview also described Claude Managed Agents' approach of separating sessions and the harness from the sandbox that executes code, with credentials mounted separately through a proxy so that a sandbox failure does not take down the entire session; Dreaming, meanwhile, is implemented as one managed agent inspecting another agent's history and writing back to memory. The team said that as safety work for model releases has grown heavier, recent efforts have shifted more toward safety, with progress on features such as skills and memory slowing relatively.
In a Latent Space podcast interview, Anthropic engineering lead Katelyn Lesse and product lead Angela Jiang discussed the Opus 5.5 release, covering adjustments to the model's writing style, pricing based on intelligence per dollar, and API changes such as requiring thinking blocks to be retained across turns to prevent distillation. The interview also described Claude Managed Agents' approach of separating sessions and the harness from the sandbox that executes code, with credentials mounted se
a16z partners Ben Horowitz and Michael McGuiness wrote the article, using Databricks' four generations of technological evolution—from Spark and Lakehouse to Lakebase and agents—as a thread to discuss how agents are changing enterprises. The article states that Databricks has surpassed $7 billion in annualized revenue, up 80% year over year, at a valuation of $190 billion, with more than 17 million databases launched by agents every day. The authors argue that enterprises need an agent-facing system of record that captures definitions, permissions, and evaluation results, and believe that in the future more people's work will shift toward specifying tasks, delegating execution, and checking results.
a16z partners Ben Horowitz and Michael McGuiness wrote the article, using Databricks' four generations of technological evolution—from Spark and Lakehouse to Lakebase and agents—as a thread to discuss how agents are changing enterprises. The article states that Databricks has surpassed $7 billion in annualized revenue, up 80% year over year, at a valuation of $190 billion, with more than 17 million databases launched by agents every day. The authors argue that enterprises need an agent-facing sy
In episode 20 of the Interconnects podcast, Nathan Lambert and Florian Brand discussed how Kimi K3.1 Pro ranks high on Artificial Analysis but receives mixed reviews, arguing that its training leans more toward exploitation than exploration, with weaker performance on long-horizon benchmarks such as ProgramBench. The two also discussed the cybersecurity controversy surrounding open-weight models, noting that no large-scale exploitation occurred after the disclosure of the PHP2Shell vulnerability, and talked about Anthropic not hiring junior employees as well as Florian's experience using Muse to handle everyday tasks such as shopping.
In episode 20 of the Interconnects podcast, Nathan Lambert and Florian Brand discussed how Kimi K3.1 Pro ranks high on Artificial Analysis but receives mixed reviews, arguing that its training leans more toward exploitation than exploration, with weaker performance on long-horizon benchmarks such as ProgramBench. The two also discussed the cybersecurity controversy surrounding open-weight models, noting that no large-scale exploitation occurred after the disclosure of the PHP2Shell vulnerability
At CASPA's 35th Annual Conference, SemiAnalysis founder Dylan Patel and Bernstein analyst Stacy Rasgon discussed the outlook for AI infrastructure investment, arguing that with the U.S. 10-year Treasury yield breaking above 5% and unsecured bonds from emerging cloud service providers yielding close to 15%, financing capacity is becoming a key constraint on expansion. The two judged that substrates and memory are bottlenecks harder to scale than power, that HBM requires roughly 3 to 4 times as many wafers as DDR5, and that if new capacity is concentrated in releases after 2028, memory prices and profit margins could come under pressure. Nvidia still supports its pricing through performance gains, but Google TPU, Amazon Trainium, and the custom chips of OpenAI and Anthropic are eroding its share and bargaining power.
At CASPA's 35th Annual Conference, SemiAnalysis founder Dylan Patel and Bernstein analyst Stacy Rasgon discussed the outlook for AI infrastructure investment, arguing that with the U.S. 10-year Treasury yield breaking above 5% and unsecured bonds from emerging cloud service providers yielding close to 15%, financing capacity is becoming a key constraint on expansion. The two judged that substrates and memory are bottlenecks harder to scale than power, that HBM requires roughly 3 to 4 times as ma
OpenAI published 722 mathematics papers at once, covering number theory, geometry, and theoretical computer science, claiming progress on problems such as the Riemann Hypothesis, Catalan's constant, and the Kakeya conjecture, without peer review. Fields Medalist Hugo Duminil-Copin said 16 problems related to his research had been flattened by AI, and some doctoral students face their topics being covered and their degrees affected. The article also mentions an Anthropic researcher claiming to have disproved the Jacobian conjecture, and Shing-Tung Yau leading the establishment of a mathematics and artificial intelligence laboratory alliance.
OpenAI published 722 mathematics papers at once, covering number theory, geometry, and theoretical computer science, claiming progress on problems such as the Riemann Hypothesis, Catalan's constant, and the Kakeya conjecture, without peer review. Fields Medalist Hugo Duminil-Copin said 16 problems related to his research had been flattened by AI, and some doctoral students face their topics being covered and their degrees affected. The article also mentions an Anthropic researcher claiming to ha
Moonshot AI released Kimi K3.1 Pro on October 24, scoring 65.7% on Terminal-Bench 4.0, but users of the 199-yuan and 699-yuan plans complain that quotas are insufficient, and some developers estimate that the API value of the old plans is about 1.94 times that of the new plans. K3.1 Pro's per-million-Token input and output prices dropped from K3's 20 yuan and 100 yuan to 13 yuan and 68 yuan, while complex Agent tasks can consume millions of Tokens. A CAICT report shows that in June 2026, domestic daily average Tokens were about 500 trillion, nearly four times the figure in December 2025; the author believes that the compute shortage will still constrain independent model companies over the long term.
Moonshot AI released Kimi K3.1 Pro on October 24, scoring 65.7% on Terminal-Bench 4.0, but users of the 199-yuan and 699-yuan plans complain that quotas are insufficient, and some developers estimate that the API value of the old plans is about 1.94 times that of the new plans. K3.1 Pro's per-million-Token input and output prices dropped from K3's 20 yuan and 100 yuan to 13 yuan and 68 yuan, while complex Agent tasks can consume millions of Tokens. A CAICT report shows that in June 2026, domesti
Derek Dicker, Vice President of AMD's Enterprise Business Group, said in an interview that agentic AI has the CPU handle orchestration of tools, APIs, databases, and runtime environments, and in some deployments the CPU-to-GPU ratio has approached 1:1. AMD recently announced Helion, a rack-scale solution for agentic AI infrastructure; its CPU-only configuration is based on sixth-generation EPYC processors and supports up to 32,768 cores. Mercury Research data shows that in the second quarter of 2026, AMD accounted for 34.5% of x86 server processor shipments, while Intel accounted for 65.5%; in the data center comparison between EPYC and Xeon SP, AMD's share was 46.4%.
Derek Dicker, Vice President of AMD's Enterprise Business Group, said in an interview that agentic AI has the CPU handle orchestration of tools, APIs, databases, and runtime environments, and in some deployments the CPU-to-GPU ratio has approached 1:1. AMD recently announced Helion, a rack-scale solution for agentic AI infrastructure; its CPU-only configuration is based on sixth-generation EPYC processors and supports up to 32,768 cores. Mercury Research data shows that in the second quarter of
Meta Q3 revenue was $64.6 billion, up 26% year-over-year, of which advertising contributed $63.06 billion, accounting for 97.6% of total revenue. The personal agent Muse, launched on September 8, surpassed 10 million downloads within 40 days of launch, but high capital expenditure turned free cash flow negative at $2.12 billion, and the full-year Capex guidance floor was raised from $130 billion to $135 billion. Muse currently mainly relies on subscription fees, with shopping being the most frequent use case; more than 25% of users have used the shopping feature, with average spend per transaction exceeding $200.
Meta Q3 revenue was $64.6 billion, up 26% year-over-year, of which advertising contributed $63.06 billion, accounting for 97.6% of total revenue. The personal agent Muse, launched on September 8, surpassed 10 million downloads within 40 days of launch, but high capital expenditure turned free cash flow negative at $2.12 billion, and the full-year Capex guidance floor was raised from $130 billion to $135 billion. Muse currently mainly relies on subscription fees, with shopping being the most freq
Google CEO Pichai revealed in the "Sources" interview that Gemini 4 Argon processed about 25 trillion tokens in 72 hours of launch, 3 times that of Gemini 3.0 in the same period, and is still rate-limited due to compute constraints. He explained for the first time the reason for cutting Gemini 3.5 Pro: the team found a gap in the model on Agentic Coding, first used the Flash version to complete intensive iterations from 3.5 to 3.8 in about 3 months, while Argon pretraining had already been completed at that time, so they injected the accumulated gains into Argon post-training. He also mentioned that Google internally has entered the RSI stage of researchers and models collaborating, and said the next-generation Gemini 5.0 will be much stronger than Argon.
Google CEO Pichai revealed in the "Sources" interview that Gemini 4 Argon processed about 25 trillion tokens in 72 hours of launch, 3 times that of Gemini 3.0 in the same period, and is still rate-limited due to compute constraints. He explained for the first time the reason for cutting Gemini 3.5 Pro: the team found a gap in the model on Agentic Coding, first used the Flash version to complete intensive iterations from 3.5 to 3.8 in about 3 months, while Argon pretraining had already been compl
Moonshot AI released Kimi K3.1 Pro on October 24, retaining K3's 2.8 trillion parameters without redoing pretraining, and raised its Artificial Analysis composite intelligence index from about 44 to about 55 by increasing post-training compute and joint reinforcement learning across hundreds of domains. Its pricing is $2 and $10 per million input/output tokens, or $12 for 1 million tokens each, 16 times the $0.75 idle-time rate of MiniMax M3.1 Flash and DeepSeek V4.1 Flash-1017. The article also compares the three vendors' respective reinforcement in 3D modeling and financial research, full project development, and computer operation and cross-application workflows, and notes that some users report 3D tasks consume tokens so fast that they hit the quota limit.
Moonshot AI released Kimi K3.1 Pro on October 24, retaining K3's 2.8 trillion parameters without redoing pretraining, and raised its Artificial Analysis composite intelligence index from about 44 to about 55 by increasing post-training compute and joint reinforcement learning across hundreds of domains. Its pricing is $2 and $10 per million input/output tokens, or $12 for 1 million tokens each, 16 times the $0.75 idle-time rate of MiniMax M3.1 Flash and DeepSeek V4.1 Flash-1017. The article also
gihyo.jp reviews multiple releases and reports from late October, organizing AI-era defense challenges along two lines: vulnerability remediation implementation and permission management. Anthropic Deputy CISO Jason D. Clinton says that within about 18 months, high cyber capability will spread to small, cheap models, but fixes still need to be applied to running systems. The article cites OpenAI's discovery of Brotli's CVE-2026-111497 during testing in an isolated environment, and introduces remediation and credential control solutions such as AWS Continuum, Google Agent Compute Engine, and 1Password for Agents, as well as Microsoft's report on Storm-0940 using AI to automate attacks.
gihyo.jp reviews multiple releases and reports from late October, organizing AI-era defense challenges along two lines: vulnerability remediation implementation and permission management. Anthropic Deputy CISO Jason D. Clinton says that within about 18 months, high cyber capability will spread to small, cheap models, but fixes still need to be applied to running systems. The article cites OpenAI's discovery of Brotli's CVE-2026-111497 during testing in an isolated environment, and introduces rem
Huxiu reviews the global AIDC construction wave, noting that combined computing power investment is expected to reach $830 billion in 2026, up 79% year over year, but as of July 2026 more than 95% of NVIDIA's Grace Blackwell capacity has not yet been put into use. Industry consensus is that about 60% or more of giant capital expenditure corresponds to clear customers, while the remaining 30-40% is preemptive capacity grabbing. The article cites cases such as CoreWeave's Q2 revenue of $2.575 billion and net loss of $626 million to show that the three conditions—single-park revenue covering costs, stable utilization and rental prices, and follow-on financing continuing—are currently not broadly verified, and the certain winner may only be equipment seller NVIDIA.
Huxiu reviews the global AIDC construction wave, noting that combined computing power investment is expected to reach $830 billion in 2026, up 79% year over year, but as of July 2026 more than 95% of NVIDIA's Grace Blackwell capacity has not yet been put into use. Industry consensus is that about 60% or more of giant capital expenditure corresponds to clear customers, while the remaining 30-40% is preemptive capacity grabbing. The article cites cases such as CoreWeave's Q2 revenue of $2.575 bill