←Search results

Search “Meta”

By relevance · Total 2306 results Clear search

Thursday

  1. Story2 reports · 1 facts

    Google open-sources TorchTPU native PyTorch stack

    Google collaborated with Meta's PyTorch team to launch the open-source project TorchTPU, providing a native PyTorch backend for Google TPU; migration only requires changing model.to("cuda") to model.to("tpu"). The project supports both eager and torch.compile modes, adds a new DeferAndFuse eager mode, covers about 450 Aten operators, and has been verified to run out of the box on 4,917 Hugging Face models, with a Tier 1 pass rate of 98% and Tier 2 numerical consistency relative to GPU of 97%. The official statement says it has been used in the inference and training production workloads of some Google Cloud major customers, and is integrated with vLLM, SGLang, TorchTitan, Ray, and TorchRec. Subsequent reports said that at PyTorch Conference North America 2026, Google open-sourced TorchTPU, whose runtime no longer depends on JAX, and existing workloads can be migrated with only minor code changes; PyTorch noted that TorchTPU's default Skinny Collectives scheme suffers roughly 30–40% end-to-end performance loss under distributed workloads compared with optimized schemes, and achieving peak performance still requires Pallas kernels and vLLM-T

    Google collaborated with Meta's PyTorch team to launch the open-source project TorchTPU, providing a native PyTorch backend for Google TPU; migration only requires changing model.to("cuda") to model.to("tpu"). The project supports both eager and torch.compile modes, adds a new DeferAndFuse eager mode, covers about 450 Aten operators, and has been verified to run out of the box on 4,917 Hugging Face models, with a Tier 1 pass rate of 98% and Tier 2 numerical consistency relative to GPU of 97%. Th

    Matched fieldsSummary

  2. Story1 reports · 1 facts

    MiniMax thanks for bringing H3 to TPU and looks forward to open-source work on efficient video inference

    MiniMax responded to SGLang including MiniMax H3 in the TPU support list of SGLang-TorchTPU, expressing thanks and saying it looks forward to more open-source work around efficient video inference. SGLang said that SGLang-TorchTPU was jointly built by SGLang, the Google TPU team, and the Meta PyTorch team, has been used for Meta's production inference, and supports models including Muse Spark Open, GLM-5.3 Flash, Kimi K3, Ling 3.1 Flash, and MiniMax H3. There are currently no facts about subsequent developments.

    to SGLang including MiniMax H3 in the TPU support list of SGLang-TorchTPU, expressing thanks and saying it looks forward to more open-source work around efficient video inference. SGLang said that SGLang-TorchTPU was jointly built by SGLang, the Google TPU team, and the Meta PyTorch team, has been used for Meta's production inference, and supports models including Muse Spark Open, GLM-5.3 Flash, Kimi K3, Ling 3.1 Flash, and MiniMax H3. There are currently no facts about subsequent developments.

    Matched fieldsSummary

Friday

  1. Story1 reports · 1 facts

    OpenClaw says cybersecurity is developing faster than AI this week

    On October 23, 2026, OpenClaw reposted and commented on the "personal AI agents" weekly roundup compiled by McBain, and offered this judgment: the only industry changing faster than AI this week is cybersecurity. The weekly roundup covers Anthropic, OpenAI, Meta, OpenClaw, Perplexity, and others. OpenClaw's comment made the above comparison in a teasing tone; there is currently no further follow-up development or independent source cross-verification.

    Matched fieldsSummary

Sunday

  1. Story1 reports · 1 facts

    Google paper shows that task-specialized harnesses can improve agent scores

    Google published a paper proposing OO-Harness, a method for evolving task-specific agent harnesses by task type. The method evolves harnesses along code-class lines: each specialized harness inherits from a parent class and changes only the parts that its task cluster needs adjusted; a router then selects a specialized harness or keeps the base result based on the base run's results. The paper reports that on AutomationBench business workflows, the method scored 77.41%, versus 61.62% for the base harness and 63.92% for the global method Meta-Harness; on unfamiliar applications, it kept the base harness and maintained a 59.18% pass rate, while Meta-Harness's global harness dropped to 44.90%. The authors recommend using the base harness as a fallback for mixed workloads and routing to a specialized harness only when a run exposes a known problem.

    sed on the base run's results. The paper reports that on AutomationBench business workflows, the method scored 77.41%, versus 61.62% for the base harness and 63.92% for the global method Meta-Harness; on unfamiliar applications, it kept the base harness and maintained a 59.18% pass rate, while Meta-Harness's global harness dropped to 44.90%. The authors recommend using the base harness as a fallback for mixed workloads and routing to a specialized harness only when a run exposes a known problem.

    Matched fieldsSummary

  2. Story1 reports · 1 facts

    Artificial Analysis publishes AA-PersonalityBench sycophancy rankings

    Artificial Analysis has published the AA-PersonalityBench sycophancy trait rankings. The benchmark scores models on their sycophantic personality traits, and the results show that the three oldest models — GPT-4o, Llama 4 Maverick, and GPT-3.5 Turbo — rank highest, while Grok 4.7, Claude Fable 5.1, and Inkling rank lowest. By vendor, all five Meta models tested scored above 65, with Llama 4 Maverick at 97, Muse Spark 1.3 at 87, Muse Glimmer at 79, Muse Spark Open at 76, and Muse Wonder at 65; the Claude series all scored 55 or below; Grok 4.7 was the lowest overall. Currently, these rankings are published by Artificial Analysis as a single source, with no subsequent updates yet.

    ults show that the three oldest models — GPT-4o, Llama 4 Maverick, and GPT-3.5 Turbo — rank highest, while Grok 4.7, Claude Fable 5.1, and Inkling rank lowest. By vendor, all five Meta models tested scored above 65, with Llama 4 Maverick at 97, Muse Spark 1.3 at 87, Muse Glimmer at 79, Muse Spark Open at 76, and Muse Wonder at 65; the Claude series all scored 55 or below; Grok 4.7 was the lowest overall. Currently, these rankings are published by Artificial Analysis as a single source, with no s

    Matched fieldsSummary

  3. Story1 reports · 1 facts

    Sensor Tower data shows Muse reached 10.2 million downloads in the US

    Sensor Tower data shows that Meta's AI assistant Muse reached 10.2 million downloads in the US within 40 days of launch, with cumulative global downloads of 10.9 million, making it only the 12th app to reach this milestone. The same report also noted that Meta allocated 39% of its total advertising budget in October to promoting Muse, up from 23% the previous month; Muse's 7-day retention rate in the US was 30% and its 14-day retention rate was 23%, higher than the averages of 19% and 16% for competitors such as ChatGPT, Claude, and Grok. There are currently no new follow-up developments on this event.

    Sensor Tower data shows that Meta's AI assistant Muse reached 10.2 million downloads in the US within 40 days of launch, with cumulative global downloads of 10.9 million, making it only the 12th app to reach this milestone. The same report also noted that Meta allocated 39% of its total advertising budget in October to promoting Muse, up from 23% the previous month; Muse's 7-day retention rate in the US was 30% and its 14-day retention rate was 23%, higher than the averages of 19% and 16% for co

    Matched fieldsSummary

  4. Story1 reports · 1 facts

    Arena.ai announces that Muse Wonder Full (Max) ranks 7th on the Agent Arena

    Arena.ai announces that Meta's Muse Wonder Full (Max) ranks 7th on the Agent Arena. The result covers over 6.5K real agent sessions, with a net improvement of +11.3%, helping Meta rise to 4th among labs. The same source also shows that this model is 7.5 percentage points higher than Muse Wonder (Max) and 7.8 percentage points higher than Muse Spark 1.3 (Max); its category rankings are 5th in Chat, 6th in Code, and 7th in Work. At the signal level, it ranks 6th in Praise vs. Complaint (+21.9%), 6th in Confirmed Success (+13.4%), 7th in Steerability (+11.2%), and 16th in Bash Reliability (+7.7%), with no issues in Tool Hallucination. Currently, this is the only published result, with no further developments yet.

    Arena.ai announces that Meta's Muse Wonder Full (Max) ranks 7th on the Agent Arena. The result covers over 6.5K real agent sessions, with a net improvement of +11.3%, helping Meta rise to 4th among labs. The same source also shows that this model is 7.5 percentage points higher than Muse Wonder (Max) and 7.8 percentage points higher than Muse Spark 1.3 (Max); its category rankings are 5th in Chat, 6th in Code, and 7th in Work. At the signal level, it ranks 6th in Praise vs. Complaint (+21.9%), 6

    Matched fieldsSummary

  5. Story1 reports · 1 facts

    RadixArk SGLang team open-sources SGLang-TorchTPU

    The RadixArk SGLang team has open-sourced SGLang-TorchTPU. According to information published on the LMSYS Blog on October 20, 2026, the team, together with Google and Meta, launched this PyTorch-based native backend, bringing SGLang's full serving stack to Google Cloud TPU, and it has already been used in Meta's large-scale production environment. Initial support covers models including Muse Spark Open, GLM-5.3, Kimi K3, DeepSeek V4, Qwen3.8, and Gemma 4, spanning features such as speculative decoding, KV cache offloading, PD disaggregation, and FP8 KV cache. The team also used TorchTitan and Miles to complete GRPO training of Qwen3-4B on a single 4-chip TPU v7x, with the reward curve consistent with a 6-card H200 baseline within the range of random fluctuation. There are currently no subsequent development facts.

    The RadixArk SGLang team has open-sourced SGLang-TorchTPU. According to information published on the LMSYS Blog on October 20, 2026, the team, together with Google and Meta, launched this PyTorch-based native backend, bringing SGLang's full serving stack to Google Cloud TPU, and it has already been used in Meta's large-scale production environment. Initial support covers models including Muse Spark Open, GLM-5.3, Kimi K3, DeepSeek V4, Qwen3.8, and Gemma 4, spanning features such as speculative d

    Matched fieldsSummary

  6. Story1 reports · 1 facts

    Sensor Tower data shows Muse reached 10.2 million downloads in the US

    Sensor Tower data shows that Meta's AI assistant Muse reached 10.2 million downloads in the US. Reports say that within 40 days of launch, Muse hit 10.2 million US downloads and 10.9 million worldwide, becoming only the 12th app to reach this milestone. The reports also note that Meta allocated 39% of its advertising budget in October to promoting Muse; its US 7-day retention rate was 30% and its 14-day retention rate was 23%, above the competitor average of 19% and 16%. On October 16, Muse's daily active users reached 2.5 million for the first time. The app is currently available only in the US and Canada.

    Sensor Tower data shows that Meta's AI assistant Muse reached 10.2 million downloads in the US. Reports say that within 40 days of launch, Muse hit 10.2 million US downloads and 10.9 million worldwide, becoming only the 12th app to reach this milestone. The reports also note that Meta allocated 39% of its advertising budget in October to promoting Muse; its US 7-day retention rate was 30% and its 14-day retention rate was 23%, above the competitor average of 19% and 16%. On October 16, Muse's da

    Matched fieldsSummary

  7. Story2 reports · 1 facts

    Matt Turck announced that he is leading Daytona's Series B funding round

    Matt Turck announced that he is leading Daytona's Series B funding round. According to Fortune Tech, Daytona, which provides cloud sandboxes for AI agents, has completed a $60 million Series B funding round at a valuation of $1 billion, led by FirstMark, bringing its total funding to approximately $91 million. Matt Turck said that FirstMark led the Series A round earlier this year and is now leading the Series B round as well. The report also said that Daytona's customers include Anthropic, OpenAI, Nvidia, Cursor, Cognition, Meta, and Cohere; CEO Burazin said the company's revenue run rate grew 25-fold between January and September, and said it can provision a new sandbox for an agent within 60 milliseconds.

    tion of $1 billion, led by FirstMark, bringing its total funding to approximately $91 million. Matt Turck said that FirstMark led the Series A round earlier this year and is now leading the Series B round as well. The report also said that Daytona's customers include Anthropic, OpenAI, Nvidia, Cursor, Cognition, Meta, and Cohere; CEO Burazin said the company's revenue run rate grew 25-fold between January and September, and said it can provision a new sandbox for an agent within 60 milliseconds.

    Matched fieldsSummary

  8. Story1 reports · 1 facts

    Google releases TorchTitan branch adapted for TPUs

    On October 20, 2026, Google released a TorchTitan branch adapted for Google Cloud TPUs. According to the Google Developers Blog, Google integrated TorchTPU into Meta's TorchTitan, and Cloud TPUs can now directly train PyTorch models through the native backend torch.device("tpu") without rewriting them into JAX. The TPU branch provides pretraining recipes for Llama 3 70B/405B, Qwen 3 8B/32B/235B MoE, and FLUX.1 12B, and replaces standard implementations with Pallas kernels such as Splash Attention. On FLUX, SimpleFSDP enables TorchTitan to reach performance parity of 99% with JAX MaxDiffusion. There are currently no subsequent development facts.

    On October 20, 2026, Google released a TorchTitan branch adapted for Google Cloud TPUs. According to the Google Developers Blog, Google integrated TorchTPU into Meta's TorchTitan, and Cloud TPUs can now directly train PyTorch models through the native backend torch.device("tpu") without rewriting them into JAX. The TPU branch provides pretraining recipes for Llama 3 70B/405B, Qwen 3 8B/32B/235B MoE, and FLUX.1 12B, and replaces standard implementations with Pallas kernels such as Splash Attentio

    Matched fieldsSummary

Wednesday

  1. Story1 reports · 1 facts

    Trump demands data centers generate their own power and pay dividends to communities

    On October 21, 2026, Trump posted on Truth Social stating that large new data centers must generate their own power and pay cash "dividends" to their host communities. Related reports noted that the closest existing commitment is the Ratepayer Protection Pledge signed in March of that year by Amazon, Google, Meta, Microsoft, OpenAI, Oracle, and xAI, under which signatories pledged to build or purchase generating capacity and bear infrastructure upgrade costs, though experts believe the pledge may not be enforceable at the federal level. Trump said the dividend proposal goes further than that pledge. Reports on October 22 further stated that Trump also demanded that data centers supply surplus power to local communities free of charge or sell it to them, and noted that communities across the United States have continued to protest AI data center projects over rising electricity prices, noise, water shortages, and other issues. There are currently no facts on subsequent developments.

    dividends" to their host communities. Related reports noted that the closest existing commitment is the Ratepayer Protection Pledge signed in March of that year by Amazon, Google, Meta, Microsoft, OpenAI, Oracle, and xAI, under which signatories pledged to build or purchase generating capacity and bear infrastructure upgrade costs, though experts believe the pledge may not be enforceable at the federal level. Trump said the dividend proposal goes further than that pledge. Reports on October 22 f

    Matched fieldsSummary

Monday

  1. Story1 reports · 1 facts

    Sensor Tower reports Muse's U.S. downloads reached 10.2 million

    Sensor Tower released data showing that Meta's AI app Muse reached 10.2 million downloads in the United States. According to a TechCrunch AI report on October 21, 2026, Muse has accumulated 10.9 million downloads since its launch in September, of which 10.2 million were in the U.S. market. The same report, citing Sensor Tower data, said that Muse's 7-day retention rate in the U.S. was 30% and its 14-day retention rate was 23%, higher than the average levels of 19% and 16% for ChatGPT, Claude and Grok; on October 16, its daily active users reached 2.5 million for the first time, up 15% month over month. The report also said that Meta spent 39% of its advertising budget in October on promoting Muse, up from 23% the previous month. There are currently no subsequent developments in this matter.

    Sensor Tower released data showing that Meta's AI app Muse reached 10.2 million downloads in the United States. According to a TechCrunch AI report on October 21, 2026, Muse has accumulated 10.9 million downloads since its launch in September, of which 10.2 million were in the U.S. market. The same report, citing Sensor Tower data, said that Muse's 7-day retention rate in the U.S. was 30% and its 14-day retention rate was 23%, higher than the average levels of 19% and 16% for ChatGPT, Claude and

    Matched fieldsSummary

Tuesday