NAll news

All news

Wednesday916 items

  1. Databricks BlogFirst-party
    AI score37

    Claude Platform natively supports OpenTelemetry, with trace data integrated into Databricks MLflow

    Claude Platform's native OpenTelemetry export has entered Public Beta. Traces from the Anthropic SDK, Messages API, and Claude Managed Agents can be written to Unity Catalog tables via Zerobus Ingest and viewed in Databricks MLflow. Developers can use Analyze trace and the Beta-stage MLflow Issues to pinpoint incorrect answers and duplicate tool calls, then evaluate the effect of changes using the same set of test cases; inputs and outputs are captured only when content mode is set to full.

    Unity Catalog and Databricks MLflow

    #Product updates#Agents#English text is being prepared.

  2. Alex Heath
    AI score58

    Sound Ventures announces investment in generative media inference company fal

    Sound Ventures' Alex Heath stated that Sound Ventures has invested in generative media inference company fal, and praised its team and execution. The quoted original fal post stated that fal announced the completion of its Series E and Series F funding rounds, totaling $1.65 billion, with Series E led by Sequoia and Series F led by BessemerVP, with NVIDIA continuing to participate in the investment.

    Quoted postfal · @fal

    Today, we’re announcing we raised our Series E & F rounds for a combined total of $1.65B. Our Series E was led by @sequoia and our Series F was led by @BessemerVP with continued investment from @nvidia. We’re excited to push the boundaries of what’s possible even more next year and this funding will enable the team to build the way we want to.

    Same story · 8 reports

    #Industry news#English text is being prepared.

  3. Rohan PaulResearch
    AI score72

    Anthropic open-sources FlashProve, which can verify on an ordinary CPU which model a response came from

    Anthropic has open-sourced FlashProve, enabling servers to mathematically prove that a given response came from a specified model, with verification completable on an ordinary CPU. On Meta's Llama 4 Maverick (400B parameters), a single Nvidia H100 generates proofs at 110 tokens/second, each about 19MB, and an ordinary CPU verifies one in about 17 seconds; the only previously known system capable of proving this model, VerInf, generated 93.6GB proofs, with verification taking nearly a day. The system was built by Claude, and because the code is difficult to audit, its security guarantees are mathematically proven in Lean 4.

    Quoted postAnthropic · @AnthropicAI

    New Anthropic research: progress on provable inference for large language models. Here, we introduce FlashProve. Building on established cryptographic techniques, FlashProve allows developers to verify whether a given response came from a given AI model. It’s substantially more efficient than prior open-source methods for doing so. These proofs have a range of applications. They can, for instance, make it harder for attackers to hijack a chat, or enable safety evaluators to verify that the model they’re testing is the same one being served to customers. For our full write-up, see here: https://t.co/OIZWPb6QZa Further progress in provable inference will likely be an important part of improving trust in AI deployment. To that end, we're open-sourcing FlashProve’s code, here: https://t.co/7ky3Ir8V06 We’re grateful to our collaborators: Alessandro Chiesa, @GiacomoFenzi, @gwberg95, @miranda__christ, Rahul Dalal (https://t.co/pjWKLHWXyz) and Ben Mayer (@transmissions11).

    Same story · 14 reports

    #English text is being prepared.#Safety / alignment#Anthropic

  5. @completeskeptic
    AI score18

    Allie the Icon shares Jev multi-intent handling tips

    Allie the Icon proposes using a "shopping cart" pattern to handle multiple intents in a single user query: recursively pick the next item until Jev determines it's done. She says she has used this method to build a coffee ordering bot for her company's Slack (with the help of DoorDash's CLI), and also used it for fast multi-intent handling in her smart home system. Diogo Almeida reposted it and praised it as "hard-won experience."

    Quoted postAllie the Icon · @allietheicon

    One thing that can be tricky to figure out is how to use Jev's primitives to handle more than one intent in a single user query. My favorite way to approach this is what I call the "shopping cart" pattern: recursively pick the next thing until Jev says we're done. I've personally used this to make a coffee ordering bot for our company Slack (using @DoorDash's CLI!), as well as a fast multi-intent handler for my smart home system. Give it a shot!

    #English text is being prepared.

  6. The DecoderResearch
    AI score60

    Stanford study: 44% of 50 widely cited LLM findings have become invalid

    Stanford researchers Owen Queen and James Zou retested 50 widely cited LLM findings using 20 models released by 8 developers between July 2024 and September 2026, and found that 44% had become invalid, another 34% had weakened markedly, only 16% still held, and 6% had become stronger; among the 22 findings with an estimated time to invalidation, the average lifespan was only 15.7 months. Findings about what models cannot do became invalid fastest: of 9 reasoning-related findings, 5 had become invalid and none held or improved, involving letter counting, arithmetic in unfamiliar bases, loss of information in the middle of long contexts, and abandoning the correct answer after a user pushback. In a case study, GPT-3.5 once fabricated 55% of references, GPT-6.1 Sol 0%, and Claude Sonnet 5.5 1%, but Qwen3.7-Flash still reached 69%. The researchers recommend framing empirical LLM findings as observations on a specific model at a specific time, and publicly releasing prompts and analysis code to enable retesting.

    #Papers/Research#English text is being prepared.#English text is being prepared.

  7. The Decoder
    AI score71

    OpenAI says GPT-6 Astra designed a hardware device from scratch in under a week

    OpenAI says its hardware team used GPT-6 Astra to design a device that lets users converse with the Codex coding agent, completing it as a side project in under a week. Astra handled all stages including enclosure design, PCB layout with component selection and routing, firmware, and the user interface; a contract manufacturer produced a small number of prototypes in under a month, but because Astra did not always account for manufacturing constraints, tolerances, and assembly details, it went through several rounds of revisions. OpenAI also says an internal model with undisclosed details scored 0.875 on BenchCAD 2.0, while GPT-6 Astra scored 0.594 at its highest reasoning level; its hardware team uses Codex and ChatGPT in areas such as robotics and compute infrastructure to handle CAD models, manufacturing drawings, and simulations, with token usage up 23-fold since February.

    Same story · 10 reports

    #Industry news#Agents#English text is being prepared.

  8. Larsen Cundric
    AI score28

    Multi-agent collaboration ultimately can't avoid ticket tracking

    The author believes that every multi-agent setup will eventually reinvent a ticket-tracking system; his own approach is to put the tracking table directly into Notion. The quoted post further explains: he uses Notion as a ledger for cloud agents, with one row per PR, recording the responsible agent, CI status, and next steps; agents check the table before starting work, thereby preventing two agents from fixing the same bug.

    Quoted postLarsen Cundric · @larsencc

    I use Notion as a ledger for my @bot cloud agents. I have a bot that takes in all my PR work and hands it out to cloud agents. With several running at once, the problem stops being code. It's knowing what's in flight, who owns it, and what's stuck and why. Chat history is bad at that, really bad. It's long, it scrolls away, and every bot and agent has its own. So everything goes in one table: > one row per PR > row is created immediately when work starts > store owner cloud agent, CI status, next step > agents check the table before starting > asking "where's that PR?" just reads the row > daily sweep against the real PRs > done rows hidden, never deleted The result is that no two agents fix the same bug, and I can ask about any PR and get a straight answer. So far it's only lost 11 PRs, and those never got a row... More on this setup soon ✌️

    Same story · 2 reports

    #English text is being prepared.#Agents

  9. Karina Nguyen
    AI score23

    Karina can't stop thinking about GLM 5.2 topping the leaderboard at low cost

    Thoughtful posted that GLM 5.2 is 5x cheaper than Opus 4.8 and 11x cheaper than Fable 5, yet it tops PostTrainBench, and argued that low cost makes personalized intelligence economically viable. Karina reposted the post and wrote "can't stop thinking about it," without elaborating further.

    Quoted postThoughtful · @thoughtfullab

    GLM 5.2 is 5x cheaper than Opus 4.8 and 11x than Fable 5, yet it tops PostTrainBench. That’s exciting because lower costs make personalized intelligence economically viable. Every company and country should be able to own models trained on its own data and have sovereignty over it. The future is millions of models, each crafted around the data, values, and decisions of the people who rely on them.

    #English text is being prepared.

  10. UCSD Hao AI LabFirst-partyProduct release
    AI score70

    Hao AI Lab integrates MiniMax H3 Flex into FastVideo, delivering up to 7x inference speedup

    Hao AI Lab integrated Pika Labs' open-source MiniMax H3 Flex into FastVideo, achieving inference speeds up to 7 times that of the official open-source inference code. On a single GB200, generating a 5-second 480p video with audio using the same prompt and random seed took 59.7 seconds with the official code, 26.6 seconds with FastVideo BF16, 16.8 seconds with NVFP4 QAD, and 15.2 seconds with FP4 attention. The official code only supports a single GPU; after FastVideo added multi-GPU support, generating a 15-second video on 4 GPUs was 7 times faster; generating a 5-second 480p first-and-last-frame video on a single DGX Spark took 59.5 seconds. The project also releases the weights, inference code, and complete training recipe, and supports the 2K model of Open-H3-Regenerate-2K.

    Same story · 4 reports

    #English text is being prepared.#Video#Multimodal

  11. UCSD Hao AI LabFirst-partyProduct release
    AI score62

    Hao AI Lab integrates MiniMax H3 Flex into FastVideo, single-GPU…

    Hao AI Lab integrated Pika's open-source MiniMax H3 Flex into FastVideo. On a single GB200, generating a 5-second clip with audio takes 59.7 seconds with the official code, 26.6 seconds with FastVideo BF16, 16.8 seconds with NVFP4 QAD, and 15.2 seconds with FP4 attention. The official code only supports a single GPU, while FastVideo adds multi-GPU support, generating a 15-second clip 7 times faster on 4 GPUs. NVFP4 QAD, by distilling into BF16 Flex, comes 14-17% closer to BF16 than ordinary quantization; FP4 attention reduces a 768p 15-second clip from 101.8 seconds to 85.3 seconds. The project also supports running on DGX Spark, taking 59.5 seconds for a 5-second 480p first-and-last-frame clip, and supports the 2K model of Open-H3-Regenerate-2K, with open weights, inference code, and a complete training recipe.

    Same story · 4 reports

    #English text is being prepared.#Video#Multimodal

  12. UCSD Hao AI LabFirst-partyResearch
    AI score58

    Hao AI Lab uses NVFP4 QAD quantization to bring video generation close to BF16 precision

    Hao AI Lab adopted NVFP4 QAD in FastVideo to perform distillation quantization on BF16 Flex; the result is closer to BF16 than ordinary quantization, narrowing the gap by 14-17%. Long video clips are limited by attention computation, and FP4 attention reduces the generation time of a 768p 15-second clip from 101.8 seconds to 85.3 seconds.

    Same story · 4 reports

    #English text is being prepared.#Video#Reasoning

  13. UCSD Hao AI LabFirst-partyProduct release
    AI score69

    Hao AI Lab integrates MiniMax H3 Flex into FastVideo, single...

    Hao AI Lab has integrated Pika Labs' open-source MiniMax H3 Flex into FastVideo. The official open-source inference code takes 59.7 seconds to generate a 5-second 480p video with audio on a single GB200, while FastVideo BF16 takes 26.6 seconds, NVFP4 QAD takes 16.8 seconds, and FP4 attention takes 15.2 seconds. The team claims it is up to 7 times faster than the official code, and has added multi-GPU support, generating a 15-second clip 7 times faster on 4 GPUs. NVFP4 QAD distills the BF16 Flex model, coming 14-17% closer to BF16 than ordinary quantization; FP4 attention reduces a 768p 15-second clip from 101.8 seconds to 85.3 seconds. This clip takes 59.5 seconds on a single DGX Spark. FastVideo also supports the 2K model of Open-H3-Regenerate-2K, and the weights, inference code, and complete training recipe have all been released.

    #English text is being prepared.#Video#Reasoning

  14. UCSD Hao AI LabFirst-partyProduct release
    AI score66

    FastVideo Supports the Open-H3-Regenerate-2K Model

    Hao AI Lab announced that FastVideo now supports the Open-H3-Regenerate-2K model from HaoZhe65347 and NVIDIA AI, and provided a link to the 2K model. The series previously covered FastVideo's inference optimization of MiniMax H3 Flex, open-sourced by Pika Labs: on a single GB200, generating a 5-second 480p video with audio takes 5.8 seconds, up to 7 times faster than the official open-source inference code, with newly added multi-GPU support.

    #Product updates#Video#Open-source ecosystem

  15. UCSD Hao AI LabFirst-partyProduct release
    AI score65

    Hao AI Lab Integrates MiniMax H3 Flex into FastVideo, Single-GPU…

    Hao AI Lab announced that Pika Labs' open-source MiniMax H3 Flex has been integrated into FastVideo. The team claims its inference speed is up to 7 times faster than the official open-source inference code, generating a 5-second 480p video with audio in 5.8 seconds on a single NVIDIA GB200. Under the same prompt and random seed, the official code takes 59.7 seconds for a 5-second clip with audio, while FastVideo BF16 takes 26.6 seconds, NVFP4 QAD takes 16.8 seconds, and FP4 attention takes 15.2 seconds; the official code only supports a single GPU, whereas FastVideo adds multi-GPU support, making a 15-second clip 7 times faster on 4 GPUs. NVFP4 QAD uses BF16 Flex as its distillation target, coming 14-17% closer to BF16 than ordinary quantization; FP4 attention reduces a 768p 15-second clip from 101.8 seconds to 85.3 seconds. A 5-second 480p first-and-last-frame clip takes 59.5 seconds on a single DGX Spark, and it supports the Open-H3-Regenerate-2K model from Hao AI Lab and NVIDIA. The weights, inference code, and complete training recipe have all been released.

    Same story · 4 reports

    #Product updates#Video#Open-source ecosystem

  17. Peter Yang
    AI score25

    Peter Yang Complains About Sexbot Sharing Data Between Clients

    Luis Vale announced the launch of the companion product Sexbot with @maryamyy0, saying it aims to alleviate loneliness; it has raised $5M and pre-orders are open. Peter Yang reposted and commented "Shared between clients lol gross," expressing disgust at the sharing of data between clients.

    Quoted postLuis Vale · @luis_vale

    Today @maryamyy0 and I are coming out of stealth with Sexbot. A companion that's always there for you. Built to fix the growing loneliness in the world. We raised $5M to build it. Pre-orders are open now. Link in the comments. https://t.co/elXihoqnQR

    #English text is being prepared.

  18. Rohan PaulResearch
    AI score55

    Harvard and MIT paper finds reasoning models' chains of thought are hard to control partly because they start with "Okay"

    A new paper from Harvard and MIT finds that reasoning models control their chains of thought poorly, partly because they habitually start with "Okay." The authors had models such as Qwen3-32B reason in uppercase or number each sentence; Qwen3-32B and OLMo-3.1-32B put about 99% of the first-token probability on "Okay"; after banning "Okay" and "Alright" and adding 1 example, the average compliance rate rose by 34 to 41 percentage points compared with instructions alone. Tasks involving hidden information such as avoiding a certain word showed more unstable responses, and across 12 models this appears to be a separate capability.

    #Papers/Research#Reasoning#Safety / alignment

  19. AnthropicResearch
    AI score61

    Anthropic releases FlashProve, which can verify the source of model responses and open-sources the code

    Anthropic has announced FlashProve, a new research direction in verifiable inference. Built on existing cryptographic techniques, it lets developers verify whether a given response came from a specified AI model, and Anthropic says it is significantly more efficient than previous open-source methods. The proof can be used to make it harder for attackers to hijack conversations, and it also allows safety evaluators to confirm that the model being tested is the same as the model serving the public. Anthropic has open-sourced the FlashProve code and lists collaborators including Alessandro Chiesa.

    Same story · 14 reports

    #Papers/Research#Reasoning#Safety / alignment

  20. Sina Tech
    AI score64

    Zhiyuan Genie G2 Works 24-Hour Consecutive Shifts at a Convenience Store; Two Robots Fail to Pass Probation

    Starting November 2, six Zhiyuan Genie G2 humanoid robots worked 24-hour consecutive shifts at a normally operating convenience store in Xuhui, Shanghai. They were evaluated on criteria including on-duty hours, business completion volume, customer reviews, and task success rate. Ultimately, four passed probation; Sister Leng, who was responsible for table cleaning, and Juan Wang, who patrolled the shelves, have not yet passed probation due to minor shortcomings in on-duty initiative and job stability. The livestream also featured experience officers patrolling the venue, having the robots identify sugary and sugar-free cola, detour when their route was temporarily blocked, and complete cooked-food picking while a box was shaking. Zhiyuan says its embodied Agent framework Genie Master, as the unified scheduling brain for the six robots, in earlier testing raised the process success rate of Genie G2 on tasks such as cooked-food packaging, picking, and restocking from 28.6% on the first day to 85.7% by the fifth day.

    Same story · 6 reports

    #Industry news#Embodied AI#Agents