←Search results

By relevance · Total 2894 results Clear search

Thursday

  1. Latent Space

    Moshi CEO Anekha Sokhal: Turning market data into actionable insights for traders

    Moshi CEO Anekha Sokhal said on the Latent Space podcast that Moshi is a recommendation engine that connects market information with traders' positions, describing it as a combination of Bloomberg, Claude and Netflix. She previously traded energy derivatives at JPMorgan, and the startup idea came from the fact that while trading, the volume of information far exceeded what one person could read. She also discussed trusted sources, document extraction (using Reducto), and the trade-offs between building models in-house and buying them.

    Moshi CEO Anekha Sokhal said on the Latent Space podcast that Moshi is a recommendation engine that connects market information with traders' positions, describing it as a combination of Bloomberg, Claude and Netflix. She previously traded energy derivatives at JPMorgan, and the startup idea came from the fact that while trading, the volume of information far exceeded what one person could read. She also discussed trusted sources, document extraction (using Reducto), and the trade-offs between b

    Matched fieldsBody

Friday

  1. Tessl: Product and Engineering Blog

    When agents write code, who should get deployment permissions?

    The Tessl blog proposes that agents can write code and suggest fixes, but should not have permission to deploy to production; deployment permissions should be given to the CI/CD pipeline. The article cites a case from the Cline repository: a malicious issue title can induce its triage agent to install a package from an attacker's fork, and via a shared build cache steal an npm token to publish an unauthorized version. The author advocates treating everything an agent reads as untrustworthy, and replacing manual code review with short-lived OIDC identities, hash-locked dependencies, and an independent verifier.

    DevCon NYC

    The Tessl blog proposes that agents can write code and suggest fixes, but should not have permission to deploy to production; deployment permissions should be given to the CI/CD pipeline. The article cites a case from the Cline repository: a malicious issue title can induce its triage agent to install a package from an attacker's fork, and via a shared build cache steal an npm token to publish an unauthorized version. The author advocates treating everything an agent reads as untrustworthy, and

    Matched fieldsBody

  2. Shao Meng

    How OpenAI defines FDE: consolidating customer-customized Harness into a unified Codex

    OpenAI's FDE team was established in early 2025, with the goal of filling the deployment gap between model capabilities and customers' real-world constraints; the approach is to first get results into production, then consolidate customized solutions into reusable products. In the early days, every use case required extensive customized integration; after Codex became stronger, the team rebuilt the application on Codex, turning customized applications into plugins and skills based on Codex primitives, and self-reports that about 90% of the customized code for each project was removed. In a semiconductor customer case, about 15 FDEs covered 10+ use cases, self-reporting about 48% time savings in targeted test generation and about 22% median performance improvement in kernel optimization; in a telecom customer service voice case, the key evaluation rose from about 60% to 90%+ in about six months.

    OpenAI's FDE team was established in early 2025, with the goal of filling the deployment gap between model capabilities and customers' real-world constraints; the approach is to first get results into production, then consolidate customized solutions into reusable products. In the early days, every use case required extensive customized integration; after Codex became stronger, the team rebuilt the application on Codex, turning customized applications into plugins and skills based on Codex primi

    Matched fieldsBody

  3. Shao Meng

    SkyRL extends GLM-5.3-Flash training context to 1 million tokens

    Through four types of system optimizations, SkyRL achieves a single-sequence training length of 1 million tokens for GLM-5.3-Flash on B200, with a peak of about 144.5 GiB, roughly 64 times that of the out-of-the-box Megatron baseline. The optimizations include running DSA on sparse kernels, chunking by query and sharding the indexer by TP, recomputing full-layer activations and CPU offloading layer inputs, and removing redundant fp32 upcasting. On real book data, the 512k peak is about 122 GiB, and 1M (CP=4) is about 109 GiB; the mean absolute difference in logprob between the trainer and vLLM is about 0.04–0.07.

    Through four types of system optimizations, SkyRL achieves a single-sequence training length of 1 million tokens for GLM-5.3-Flash on B200, with a peak of about 144.5 GiB, roughly 64 times that of the out-of-the-box Megatron baseline. The optimizations include running DSA on sparse kernels, chunking by query and sharding the indexer by TP, recomputing full-layer activations and CPU offloading layer inputs, and removing redundant fp32 upcasting. On real book data, the 512k peak is about 122 GiB,

    Matched fieldsBody

  4. Rohan Paul

    Meta's new paper proposes RSD: agents can learn from unlabeled historical trajectories

    Meta's new paper proposes the RSD method, which lets AI agents review each step after knowing the task outcome and learn from historical attempts that no one labeled as right or wrong. On ScienceWorld, Qwen3-14B's success rate rose from 12.9% to 20.6%; on ALFWorld, Qwen3-8B reached 76.6%, surpassing all tested methods, including those that received rewards or standard answers. The authors recommend keeping agents' unlabeled run records (including failures) and regenerating the review notes each round, because when notes were frozen at round 3 the ALFWorld success rate stayed at 59.7%, while updating notes each round gave 67.9%.

    Meta's new paper proposes the RSD method, which lets AI agents review each step after knowing the task outcome and learn from historical attempts that no one labeled as right or wrong. On ScienceWorld, Qwen3-14B's success rate rose from 12.9% to 20.6%; on ALFWorld, Qwen3-8B reached 76.6%, surpassing all tested methods, including those that received rewards or standard answers. The authors recommend keeping agents' unlabeled run records (including failures) and regenerating the review notes each

    Matched fieldsBody

  5. 36Kr · AI

    Exiting after just 188 days, big-company AI products are accelerating their reshuffle

    36Kr sorted through the product moves of eight companies—Tencent, Alibaba, Baidu, ByteDance, and Moonshot AI, Zhipu, MiniMax, DeepSeek—across ten AI tracks over four years, finding that competition cycles have shortened, product lifecycles have grown shorter, and renaming, integration, and shutdowns have become the norm. Tencent QClaw went from public beta on March 20 to announcing service shutdown arrangements on September 24, just 188 days, with user data to be migrated to WorkBuddy; Alibaba's Qwen announced the upcoming release of its personal assistant product Qwen Go!. Alibaba and ByteDance cover all ten tracks, Tencent and Baidu each cover nine, and DeepSeek covers only two.

    36Kr sorted through the product moves of eight companies—Tencent, Alibaba, Baidu, ByteDance, and Moonshot AI, Zhipu, MiniMax, DeepSeek—across ten AI tracks over four years, finding that competition cycles have shortened, product lifecycles have grown shorter, and renaming, integration, and shutdowns have become the norm. Tencent QClaw went from public beta on March 20 to announcing service shutdown arrangements on September 24, just 188 days, with user data to be migrated to WorkBuddy; Alibaba's

    Matched fieldsBody

  6. 36Kr · AI

    Musk says superintelligent compute capacity will be in long-term shortage, SpaceX plans to expand compute to 13 gigawatts by end of 2027

    At SpaceX's Q3 2026 earnings call, Musk said superintelligent compute capacity will face a long-term shortage, with power, cooling and infrastructure constraining supply more than chips. SpaceX's superintelligence business revenue for the quarter was $7.56 billion, surpassing the connectivity business led by Starlink; compute scale rose from 0.6 gigawatts a year earlier to 1.7 gigawatts, with plans to reach 13 gigawatts by the end of 2027. Musk said paid weekly active users of Grok Bot reached 800,000 in early October, and the company also plans to build its own gas turbine blade casting and chip factories to ease bottlenecks.

    At SpaceX's Q3 2026 earnings call, Musk said superintelligent compute capacity will face a long-term shortage, with power, cooling and infrastructure constraining supply more than chips. SpaceX's superintelligence business revenue for the quarter was $7.56 billion, surpassing the connectivity business led by Starlink; compute scale rose from 0.6 gigawatts a year earlier to 1.7 gigawatts, with plans to reach 13 gigawatts by the end of 2027. Musk said paid weekly active users of Grok Bot reached 8

    Matched fieldsBody

  7. 36Kr · AI

    AI mass-generates e-commerce videos: a 5-person team produces over a thousand per day, at under 5 yuan each

    Yunying Research Society analyzes that AIGC is taking over the mass-production stage of e-commerce recommendation videos; a team of about 5 people can generate over a thousand videos in a few minutes, at under 5 yuan each, while manual shooting costs about 100 to 200 yuan per video. The specific approach is to first break out fixed blocks and variable blocks to form a template, then use low-cost models such as Grok Video and the Seedance entry version to test hooks and directions, and after getting data, switch to Veo, Sora, and the Seedance HD version to produce finished products. The article also mentions that Douyin, Xiaohongshu, and Video Accounts all tightened AIGC content review this year, and that Ocean Engine reported at the end of August that it had shut down more than 50,000 violating accounts, and holds that once materials are no longer scarce, judgment and aesthetics become the new threshold.

    Yunying Research Society analyzes that AIGC is taking over the mass-production stage of e-commerce recommendation videos; a team of about 5 people can generate over a thousand videos in a few minutes, at under 5 yuan each, while manual shooting costs about 100 to 200 yuan per video. The specific approach is to first break out fixed blocks and variable blocks to form a template, then use low-cost models such as Grok Video and the Seedance entry version to test hooks and directions, and after gett

    Matched fieldsBody

  8. 36Kr · AI

    Altman: OpenAI's internal models surpass public models in mathematical research capability; programmers are not replaced by AI but busier and earning more

    In a 45-minute conversation, Sam Altman said that OpenAI's not-yet-public models have reached a level in some mathematical research tasks that existing public models find hard to match, and he expects to find the first candidate treatment for a major disease in about two years. He said programmers have not lost their jobs because of AI; instead they are busier, creating greater value and even earning more, because the world's demand for software far exceeds expectations; he himself has not looked at a line of code for about a year. Speaking of the next-generation personal AI assistant, he believes model intelligence is already sufficient; what is lacking is product, compute, and new hardware that can continuously perceive the real environment, and he said Dot saves him about an hour every day.

    In a 45-minute conversation, Sam Altman said that OpenAI's not-yet-public models have reached a level in some mathematical research tasks that existing public models find hard to match, and he expects to find the first candidate treatment for a major disease in about two years. He said programmers have not lost their jobs because of AI; instead they are busier, creating greater value and even earning more, because the world's demand for software far exceeds expectations; he himself has not looke

    Matched fieldsBody

  9. 36Kr · AI

    A PHP high-severity vulnerability triggers global panic; volunteer Ondřej Surý single-handedly completes the critical packaging work

    Two PHP remote code execution vulnerabilities with a CVSS score of 9.2 were disclosed; information leaked before the patch was released, and Cloudflare observed hundreds of thousands of attack attempts within two hours. Germany, Finland, France, Belgium and other countries took protective measures for public sector websites, and the U.S. CISA, the EU CERT-EU, and the UK NCSC issued urgent warnings in succession. Czech volunteer Ondřej Surý, through volunteer work, packaged the PHP source code into installable software packages, enabling millions of servers to be patched quickly; the donations he received in these 24 hours were roughly equal to those of the past year.

    Two PHP remote code execution vulnerabilities with a CVSS score of 9.2 were disclosed; information leaked before the patch was released, and Cloudflare observed hundreds of thousands of attack attempts within two hours. Germany, Finland, France, Belgium and other countries took protective measures for public sector websites, and the U.S. CISA, the EU CERT-EU, and the UK NCSC issued urgent warnings in succession. Czech volunteer Ondřej Surý, through volunteer work, packaged the PHP source code in

    Matched fieldsBody

  10. 36Kr · AI

    Liang Wenfeng needs his own Tibo: DeepSeek product department integration and the Agent path

    DeepSeek has integrated the Harness with the web, App, and open platform teams into a product department, led by Cui Tianyi, and the official website has posted recruitment positions in two directions: AI product manager and Agent Harness. The article uses OpenAI Codex lead Tibo Sottiaux and the Anthropic Claude Code team as examples to explain how the execution capabilities of coding tools can be extended into a general work task entry point. DeepSeek, while open-sourcing Harness and the inference framework DeepInfer, is also facing cost pressure on caching and transmission brought by long tasks.

    DeepSeek has integrated the Harness with the web, App, and open platform teams into a product department, led by Cui Tianyi, and the official website has posted recruitment positions in two directions: AI product manager and Agent Harness. The article uses OpenAI Codex lead Tibo Sottiaux and the Anthropic Claude Code team as examples to explain how the execution capabilities of coding tools can be extended into a general work task entry point. DeepSeek, while open-sourcing Harness and the infere

    Matched fieldsBody

  11. 36Kr · AI

    Global Embodied Intelligence Industry Development Analysis Report 2026

    In the first half of 2026, global humanoid robot shipments were about 19,000 units, with Chinese manufacturers' share rising to over 97%; in 2025, global shipments were about 25,000 units, with China accounting for about 52%, and global embodied intelligence financing exceeded USD 11 billion. The report notes that the industry has entered its first year of mass production and a commercialization inflection point, but the ratio of intended orders to actual deliveries is only about 10%–20%, and commercialization remains a key bottleneck. China and the United States are positioned in a staggered manner along the value chain: the United States is strong in the "brain" and computing capital, while China is strong in the "body" and supply chain mass production.

    In the first half of 2026, global humanoid robot shipments were about 19,000 units, with Chinese manufacturers' share rising to over 97%; in 2025, global shipments were about 25,000 units, with China accounting for about 52%, and global embodied intelligence financing exceeded USD 11 billion. The report notes that the industry has entered its first year of mass production and a commercialization inflection point, but the ratio of intended orders to actual deliveries is only about 10%–20%, and co

    Matched fieldsBody

  12. TMTPost · AGI

    The top three AI 3D players raised over 8 billion yuan in half a year; Meshy's ARR exceeds $100 million

    The three AI 3D companies VAST, Meshy, and Yingmou raised a combined total of over 8 billion yuan in half a year, among which VAST raised over 5 billion yuan in half a year with a valuation of $2 billion, and Meshy completed a $400 million funding round with a post-money valuation of $1.5 billion. As of the end of September, Meshy's ARR exceeded $100 million, and the ARR of VAST and Yingmou also reached tens of millions of dollars. The three have different roadmaps: VAST wants to build a 3D version of Douyin and open-sourced the real-time renderer EdenLive, Meshy adopts subscriptions plus pay-per-use, and Yingmou targets professional users and enterprise production workflows.

    The three AI 3D companies VAST, Meshy, and Yingmou raised a combined total of over 8 billion yuan in half a year, among which VAST raised over 5 billion yuan in half a year with a valuation of $2 billion, and Meshy completed a $400 million funding round with a post-money valuation of $1.5 billion. As of the end of September, Meshy's ARR exceeded $100 million, and the ARR of VAST and Yingmou also reached tens of millions of dollars. The three have different roadmaps: VAST wants to build a 3D vers

    Matched fieldsBody

  13. TMTPost · AGI

    Meta Q3 earnings: advertising props up 97% of revenue, Muse still burning money

    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

    Matched fieldsBody

  14. TMTPost · AGI

    Pichai Reviews Gemini 4 Argon: Why 3.5 Pro Was Cut

    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

    Matched fieldsBody

  15. Tech.eu

    Brain Fog, Misdiagnosis and Career Loss: The Menopause Cognitive Health Gap Cerebella Wants to Fill

    UK startup Cerebella has launched Cerebella Foundations, a self-guided program developed by cognitive neuroscientists to help women understand and maintain cognitive health during the menopause transition. The program is a structured course consisting of ten sessions over five weeks, each lasting about 25 to 30 minutes. A Cerebella survey of about 1,300 to 1,400 UK women found that brain fog is associated with reduced confidence, lower productivity and intention to leave work; in a study of more than 14,000 women, about 40–80% reported self-reported cognitive symptoms during the transition.

    UK startup Cerebella has launched Cerebella Foundations, a self-guided program developed by cognitive neuroscientists to help women understand and maintain cognitive health during the menopause transition. The program is a structured course consisting of ten sessions over five weeks, each lasting about 25 to 30 minutes. A Cerebella survey of about 1,300 to 1,400 UK women found that brain fog is associated with reduced confidence, lower productivity and intention to leave work; in a study of more

    Matched fieldsBody

  16. Tech.eu

    European companies urged to participate more actively in open source software development

    At the Open Source Summit Europe hosted by the Linux Foundation in Prague, Thierry Carrez, General Manager of Linux Foundation Europe, and others called on European companies to not only use open source software but also contribute upstream and participate in open source governance, in order to reduce dependence on a single vendor or country. A Linux survey shows that more than 90% of European organizations regard digital sovereignty as a key priority in their technology strategy; companies with mature open source governance have an average return on investment of more than 4x, compared with 3.6x for those without a formal governance structure. GitHub data from October shows that currently one in every three pull requests involves an AI agent, compared with less than one in ten a year ago.

    At the Open Source Summit Europe hosted by the Linux Foundation in Prague, Thierry Carrez, General Manager of Linux Foundation Europe, and others called on European companies to not only use open source software but also contribute upstream and participate in open source governance, in order to reduce dependence on a single vendor or country. A Linux survey shows that more than 90% of European organizations regard digital sovereignty as a key priority in their technology strategy; companies with

    Matched fieldsBody

  17. Semiconductor Industry Observer

    MSI releases EdgeMesa N AI+ mini PC, 1.3L chassis runs 120B large model locally

    MSI has released the EdgeMesa N AI+ mini PC, with a volume of 1.3 liters, equipped with the NVIDIA RTX Spark N1X superchip and up to 128GB of LPDDR5X unified memory, with running 120B-parameter large models locally as its primary design goal. The chip integrates a 20-core Grace CPU and a Blackwell GPU with 6144 CUDA cores; the unified memory pool lets model weights and KV cache share the same physical memory. Referring to measured results from the DGX Spark on the same platform, generation speed for 120B-class models is about 15 to 59 tokens/s, limited by 273GB/s memory bandwidth; official pricing has not yet been announced, and the article estimates it may fall in the 30,000 to 40,000 RMB range.

    MSI has released the EdgeMesa N AI+ mini PC, with a volume of 1.3 liters, equipped with the NVIDIA RTX Spark N1X superchip and up to 128GB of LPDDR5X unified memory, with running 120B-parameter large models locally as its primary design goal. The chip integrates a 20-core Grace CPU and a Blackwell GPU with 6144 CUDA cores; the unified memory pool lets model weights and KV cache share the same physical memory. Referring to measured results from the DGX Spark on the same platform, generation speed

    Matched fieldsBody

  18. English text is being prepared.

    Mengtong Intelligent displays a 39-gram binocular full-color waveguide AI glasses, adopting an MCU+DSP dual-chip architecture

    Mengtong Intelligent exhibited at the Hong Kong Autumn Electronics Show a binocular full-color waveguide AI glasses weighing only 39 grams for the whole device, which the company officially calls the lightest product of its kind in the industry. The product adopts an MCU+DSP dual-chip architecture, striking a balance between low-power control and real-time task processing, and supports functions such as AI Q&A, real-time translation, speech prompting, and navigation, but the battery capacity and actual battery life have not yet been announced.

    Mengtong Intelligent exhibited at the Hong Kong Autumn Electronics Show a binocular full-color waveguide AI glasses weighing only 39 grams for the whole device, which the company officially calls the lightest product of its kind in the industry. The product adopts an MCU+DSP dual-chip architecture, striking a balance between low-power control and real-time task processing, and supports functions such as AI Q&A, real-time translation, speech prompting, and navigation, but the battery capacity and

    Matched fieldsBody

  19. sspai

    Morning Brief: Insta360 releases Insta360 Mic, Anthropic launches Claude Duet, and more

    Anthropic launches Claude Duet, an AI agent app that can work proactively, built on Claude Opus 5.5, able to call more than 6,000 connectors and plugins in Claude Marketplace; the research preview's first batch opens a waitlist to Pro and Max subscribers in the United States. Insta360 releases the wireless microphone Insta360 Mic, priced at 439 yuan for a single transmitter, with a maximum transmission distance of 400 meters and 13 hours of transmitter battery life. EA closes BioWare studio and cancels the next-generation Mass Effect title still in early development, reportedly cutting 400 positions.

    Anthropic launches Claude Duet, an AI agent app that can work proactively, built on Claude Opus 5.5, able to call more than 6,000 connectors and plugins in Claude Marketplace; the research preview's first batch opens a waitlist to Pro and Max subscribers in the United States. Insta360 releases the wireless microphone Insta360 Mic, priced at 439 yuan for a single transmitter, with a maximum transmission distance of 400 meters and 13 hours of transmitter battery life. EA closes BioWare studio and

    Matched fieldsBody