Not Everything Needs an LLM
1 tiếng nói · chưa gắn nhãn trưởng lab · 1 nền tảng · 261 tương tác đã đo · Nguồn crawl
Chi tiết nguồn
- Not Everything Needs an LLM HN: LLM · 18 thg 11, 2026
Chủ đề cộng đồng đang thảo luận
Crawl live-1795148100000 · 11:15 20 thg 11, 2026 · Xuất bản 11:32 20 thg 11, 2026
1 tiếng nói · chưa gắn nhãn trưởng lab · 1 nền tảng · 261 tương tác đã đo · Nguồn crawl
2 tiếng nói · chưa gắn nhãn trưởng lab · 2 nền tảng · chưa có số tương tác · Nguồn crawl
<img src="https://storage.googleapis.com/gweb-uniblog-publish-prod/images/Hero_2784_x_1566.max-600x600.format-webp.webp">Tasks in Gemini is now free of charge in six countries, with new features starting in the U.S. and deeper Google Workspace integration to get things done on your behalf.
1 tiếng nói · chưa gắn nhãn trưởng lab · 1 nền tảng · 62 tương tác đã đo · Nguồn crawl
Hi HN, Sanchit here from RunAnywhere (YC W26). We just open-sourced TrnRT: Trainium RunTime, our inference engine for AWS Trainium. We started 3 weeks back, and currently we’re able to serve and run GLM 5.3 flash (FP8) on AWS trn2.48xlarge, and will bring more models support soon. It’s open to contributions :) What it does today: - Serves 2 instances of GLM-5.3 flash on a single box, 8 trn2 with a cache aware router. - Can process 3.57M TPM (tokens per minute) on SemiAnalysis's AgentX benchmarks. - Runs decode as one NKI kernel per layer half, or for the 11 sparse-attention layers fuses both blocks into a single kernel, for faster performance. We started working on Trn2 because we wanted to explore it and see if we can come close to beating it’s performance in terms of cost efficiency compared to H100/H200. A big MoE with 1M context window such as GLM-5.3 flash requires a lot of memory, and memory/tb on a trn2.48xlarge costs less than half compared to 8x h200 on AWS, and hence we wanted to experiment with it and now we’re sharing some early results. But AWS's own stack didn't support the model: vLLM Neuron's latest release did not register any GLM model, and NxDI, AWS's own distributed inference framework, is no longer under active development. We also saw the drop of SGLang Neuron implementation yesterday, but there is no good support for Trn2, which is the only chip available for public use right now. We ran semiAnalysis's AgentX benchmarks, which replays real coding agent sessions: - 1 box served 28.9 RPM(requests per minute) with 32 agents at batch 12 (also known as concurrency) -The P50 TTFT with a cached prompt at 16 agents is ~1.24 seconds with a decode of 35.9 tps. Initially, the recurrence in one kda (linear attention) layer was ~2% of the layer's attention time at batch 1, and the rest was projections, transposes and XLA glue between small kernels. So we wrote the megakernel so that it covers the whole layer in 2 kernel dispatch, which made decode 25% faster at batch 8. The sparse attention kernel keep a layer’s working weights in SBUF, distributing weighted streaming across the cores and handling the interchip communication within There’s an approximation in the sparse-attention prefill, affecting about 1.2% of block selections. You can refer to the write-up for the profiling results: https://runanywhere.ai/blog/glm-5-3-flash-trainium2-nki-mega... Please give it a try yourself on a trn2.48xlarge and reproduce the AgentX results by following the steps in docs/RESULTS.md. Happy to collaborate and hear any feedback. Deploying high performant dedicated inference at lab like performance and half the price for open models such as the GLM-5.3 family, Kimi K3, Deepseek flash etc. is how we make money. You can talk to us regarding the same at founders@runanywhere.ai. We hope you find this useful. If you are working with Trainium or have experience with NKI kernels, we'd love to get your honest feedback. Feel free to share the models you want support for next, and we welcome any contributions to the repository.
3 tiếng nói · chưa gắn nhãn trưởng lab · 2 nền tảng · chưa có số tương tác · Nguồn crawl
Cloudflare has been one of the essential infrastructure companies of the web for the past decade and a half. But now, it’s turning into a neolab — at least according to a presentation at AI Engineer Code Summit by Michelle Chen , who runs Cloudflare’s AI Platform on the product side. We’ll get into what exactly a neolab is in a minute (the short version: they build models and they’re not called OpenAI or Anthropic). Cloudflare is claiming to be one because it too has begun to launch models. Two System One models , to be precise: Clef and Clairic. As Chen explained in her talk, Clef is a decision model directly inspired by TypeSafe AI’s revolutionary Jev model , while Clairic is a new type of system one model that Cloudflare calls an “extract” model . In an interview with Latent Space after her presentation, Chen hinted that Cloudflare won’t just be producing system one models. There will be large language models, too. “We're not being a neolab [to] only launch system one models, even though it makes sense for our infrastructure. We've kind of entered this Pandora's box now, right? We opened it, now the possibilities are endless.” Chen said it’s “not just system one models, but potentially reasoning models that we've been experimenting with, and other LLMs and whatnot.” However, she added that these probably won’t be pre-trained models built from scratch, like the frontier labs produce. “I don't think that's our focus, but I think we can take existing resources out there and make [them] a lot more efficient and effective.” She noted that “most of our team really just does post-training,” so that will be the focus in model development. What is a neolab and is Cloudflare really one? The term neolab literally means a ‘new lab’ and it was coined last November by Stephanie Palazzolo from The Information, who defined it as “a new generation of research-focused AI startups.” Deedy Das , a partner at Menlo Ventures and an AI influencer, runs a big list of neolabs . It includes companies like Thinking Machines Lab, Poolside, and TypeSafe AI itself. According to Das, neolabs employ elite researchers, spend a lot of compute (“usually to train its own AI model”), don't have significant revenue traction, and they’re typically startups that have raised a lot of money. Deedy Das’ list of neolabs — note that TypeSafe is among the top 20. Cloudflare doesn’t tick at least half of those boxes, since it’s a public company with significant revenue traction . In fact, earlier this month Cloudflare announced record earnings . “The third quarter was Cloudflare's best quarter ever,” said CEO Matthew Prince . “We crossed $3 billion in annualized revenue with revenue of $782 million, up 39% year over year.” So Cloudflare as a “neolab” is pushing the accepted definition. However, as Chen argued in her presentation, building models is part of a natural progression for the company : from infrastructure, to compute (e.g. the Workers platform, including now for agents), to storage, to serving open weight models like Kimi, GLM and DeepSeek (on Workers AI), to now building its own models. Jev as the “east wind” Chen readily admits that Jev inspired Cloudflare’s pivot to building models. Referencing a Chinese proverb, she framed Jev’s arrival in September as “the east wind” that unexpectedly changed how Cloudflare viewed AI models. “And so when the system one and Jev paradigm started coming on board,” Chen said in her presentation, “we were like, this makes so much sense for us to adopt, right? We don't have to send every request to an LLM. We can send some of it to a System One model.” In our conversation, Chen commented on how it took only two weeks for them to build Clef — which is a copy of Jev — whereas it took TypeSafe AI two years to build Jev. She meant that as a compliment for how hard TypeSafe AI worked to create this new type of model. “All credit to TypeSafe, really, for actually attacking this space and inventing something new here , that people could relate to so strongly and be able to use it. And, yeah, we came up with a model super quickly, but they put in a lot of legwork and we knew exactly what the API schema should look like and whatnot.” Cloudflare is, at its core, still an infrastructure company — speed and low latency is its meat and potatoes. The main benefit of System One models, Chen remarked, is that they allow more work to be done by smaller, faster models , where output is measured in milliseconds rather than seconds. Clairic and what extraction models do While Clef was a clone of Jev, Cloudflare soon came up with an original twist on System One with Clairic , launched last month as “the first model in a new class: System One extraction models.” “What Clairic does is it can take unstructured input, and it can extract, compute, estimate and forecast values from those inputs,” Chen explained in her presentation. In our interview, Chen said that her team wanted to use System One principles to find other areas where LLMs were “overkill” and didn’t always return the right answer. One of the problems they settled on was that LLMs don’t always deal well with unstructured input. Because LLMs are autoregressive, Chen explained, they are prone to hallucinate and they sometimes return malformed JSON. “So we used the same System One principles [as Jev/Clef] there, and decided on extraction.” Some examples of system one use cases at Cloudflare. She used the example of an invoice, where an LLM will sometimes get an important detail wrong — like the order number. “When you send it [the order] to an LLM, they're like, wait, I think the order number's three, maybe it's four, maybe it's five. But it's like, the number's right there, you just have to read it! ” So the idea with Clairic is that it will extract a typed value — like a number — from unstructured data, in one forward pass. Why Cloudflare is in the lab now Using the label “neolab” sounds a bit like a marketing gimmick, but it also makes a ton of sense
1 tiếng nói · chưa gắn nhãn trưởng lab · 1 nền tảng · 40 tương tác đã đo · Nguồn crawl
Hey HN, we're Yarn (W24) and we make a video editing app for startups, and we built Quebec ( https://quebec.yarn.so ), an API that gives you accurate word timestamps from speech. We needed it because cutting a word even 50ms off from its true timing can ruin talking head footage (examples in the write up). Accurate word timestamps usually come from a forced aligner, which takes the speech audio and a transcript and works out exactly when each word starts and ends. We weren't satisfied with the accuracy of Montreal Forced Aligner (MFA) or new models like ElevenLabs Cue ( https://elevenlabs.io/blog/cue-aligner ). Giving LLMs access to a waveform/spectrogram didn't help much either. So we built our own. There are two particularly interesting parts: - We wrote our own decoder around MFA's acoustic model that only lets each word land within 200ms of the rough timings. That made search small enough to solve exactly, so we don't need to use MFA's beam search. - We added a small CNN, built on the first two layers of Whisper Tiny, that nudges each word boundary by up to 50ms. We couldn't find enough hand-labelled speech that allowed for commercial use, so we had Gemini label our own (17k word boundaries). It took a couple of hours and cost under $200. A net trained on Gemini's labels scored the same as one trained on human labels (73.9% vs. 74%). One way to measure accuracy is % of word boundaries within 20ms of hand labels. Here are the results on Buckeye (hand-labelled American speech that we didn't train on): - Quebec 75% - MFA 61% - ElevenLabs Cue 30% If you want to try it, it's free until tomorrow night. You can point your agent at our docs at https://quebec.yarn.so , and send audio on its own or with your own transcript. After that, it costs $0.32 per hour with no signup (minimum $10 top-up). We serve it at cost. Caveats: only tested in English and slower than ElevenLabs. Let us know if Quebec works well for you, and send us any samples that work badly (you can email support at yarn dot so). Also let us know if you've built or found a better aligner. In theory we would just switch to a better aligner, although we would probably just push a bit harder to make ours better. Thanks! Jasper
1 tiếng nói · chưa gắn nhãn trưởng lab · 1 nền tảng · 529 tương tác đã đo · Nguồn crawl
1 tiếng nói · chưa gắn nhãn trưởng lab · 1 nền tảng · 76 tương tác đã đo · Nguồn crawl
11 tiếng nói · chưa gắn nhãn trưởng lab · 1 nền tảng · 14 tương tác đã đo · Nguồn crawl
Bài báo mô tả vòng lặp tiến hóa tác nhân: proposer tạo các ứng viên (prompts, skills, memories) và selection chọn những ứng viên có điểm cao hơn. Tuy vậy, họ nói rằng evolution có thể thất bại khi khởi động trên một số tác vụ vì đánh giá nhiễu không thể phân biệt đáng tin cậy giữa ứng viên tốt và xấu.
Agent evolution follows a simple loop: a proposer writes candidate artifacts (prompts, skills, memories) and selection accepts those that score higher. This loop has shown promise, yet practitioners face a puzzle: methods trade wins across papers, and evolution often fails to bootstrap on certain tasks. We trace this puzzle to the Resolution Hypothesis: agent evolution stalls not because methods fail to propose better artifacts, but because noisy evaluations cannot reliably identify them. Across seven evolution methods and six models (over 10M LLM calls), we find support for the hypothesis at both levels where evaluation scores are used. At the candidate level (choosing artifacts), helpful candidate artifacts exist, but selection is no better than random at telling them from harmful ones, so errors accumulate across rounds. At the method level (ranking methods), typical margins of a few percentage points are smaller than the difference between two independent runs of the same method, so published rankings often rest on luck. Together, these findings indicate that agent evolution is often bottlenecked by reliably identifying improvements. This calls for a shift in how we develop and evaluate self-improving agents, from designing ever more candidate improvement systems to making true improvements detectable and accumulable.
We trace this puzzle to the Resolution Hypothesis: agent evolution stalls not because methods fail to propose better artifacts, but because noisy evaluations cannot reliably identify them.
We trace this puzzle to the Resolution Hypothesis: agent evolution stalls not because methods fail to propose better artifacts, but because noisy evaluations cannot reliably identify them.
Across seven evolution methods and six models (over 10M LLM calls), we find support for the hypothesis at both levels where evaluation scores are used.
Together, these findings indicate that agent evolution is often bottlenecked by reliably identifying improvements. This calls for a shift in how we develop and evaluate self-improving agents, from designing ever more candidate improvement systems to making true improvements detectable and accumulable.
7 tiếng nói · chưa gắn nhãn trưởng lab · 1 nền tảng · 9 tương tác đã đo · Nguồn crawl
Latent diffusion models rely on a decoder to map denoised latents to pixels, adding complexity and inference cost. We introduce Decoding Without a Decoder (DWD), which lets a finetuned diffusion transformer produce final pixels directly through a linear head. Unlike previous methods, where the denoising output leads to final pixels, DWD separates the two roles: the original denoising head continues to drive sampling, then the new DWD head predicts the pixels at the final step. We extend pretrained latent models with DWD efficiently by finetuning only the last few transformer blocks, training the DWD head with pixel quality losses while anchoring the denoising head to the pretrained model to preserve the sample trajectory. Across latent image and video generators, DWD removes the decoder and reduces inference cost while maintaining or even improving pixel quality. On Wan video generation, it visibly reduces detail distortion and flickering, lowering warping error by 23%. Remarkably, the same separation also improves pixel diffusion models that do not even use a decoder: a final-step DWD head produces substantially better texture detail in text-to-image generation, and improves ImageNet FID from 1.55 to 1.43, a new state of the art among pixel flow models.
6 tiếng nói · chưa gắn nhãn trưởng lab · 1 nền tảng · 3 tương tác đã đo · Nguồn crawl
Deep models often rely on spurious correlations that fail on underrepresented groups. Existing methods either correct human-specified bias attributes or infer unknown biases from the learned representation space. However, annotated attributes may differ from the shortcuts actually used by a model, while a single representation provides incomplete and potentially noisy evidence. Our key insight is that a bias obscured in one feature space may be clearer in another. In this work, we propose CHORD (Cross-view Harmonic Outlier Reversal for Debiasing), a multi-view framework that allows samples to "sing in chorus" across diverse representation spaces to achieve robust cross-view consensus. CHORD aggregates aggregates within-class partitions in each view, which is a proxy for specifying candidate biases, to debias the target representation and classifier without bias supervision. Experiments on seven benchmarks across vision, language, speech, and medical imaging show consistent improvements in worst-group accuracy. Code and checkpoints are available at https://github.com/davelee-uestc/chord_debiasing_public .
4 tiếng nói · chưa gắn nhãn trưởng lab · 1 nền tảng · 2 tương tác đã đo · Nguồn crawl
The goal of this paper is to reduce the input token cost of Omni-modal large language models (Omni-LLMs) at inference time. Omni-LLMs reason jointly over audio, video and text, but the cost of the three streams is highly unbalanced: visual tokens account for the vast majority of the input, and are highly redundant. In this paper, we propose ReMo, a training-free framework that compresses visual tokens by redistributing their information across modalities: a visual token is kept only if its information appears nowhere else. ReMo achieves this in two ways: (i) it aligns audio and video in a shared space via closed-form calibration, and removes visual tokens already explained by the audio or by other visual tokens; and (ii) it replaces object-level visual tokens, identified by a lightweight object detector, with compact text proxies, short descriptions of each object and its location, conveying the same content in far fewer tokens. On Qwen2.5-Omni at two model scales, ReMo removes 54% of the input tokens with comparable accuracy. It performs comparably to the full-token model, reaching 101.2% and 101.4% of its average accuracy over five audio-visual benchmarks.
4 tiếng nói · chưa gắn nhãn trưởng lab · 1 nền tảng · 2 tương tác đã đo · Nguồn crawl
Bài báo nói rằng các câu hỏi chi tiết theo thời gian trong video dài có thể yêu cầu tìm các khoảnh khắc ngắn rải rác trong nhiều giờ. Việc đánh giá và học cho các mô hình tác nhân quyết định khi nào và nơi nào cần quan sát là khó vì bằng chứng thưa thớt có thể không đảm bảo đếm đúng hoặc không đủ để kiểm chứng toàn bộ dòng thời gian. Nhóm giới thiệu NexSight để xây dựng các tác vụ có thể kiểm chứng nhằm vừa đánh giá vừa huấn luyện hiểu video dài theo dạng tác nhân.
Fine-grained questions about long videos ask how many times an action occurs, in what order events unfold, or exactly when a state changes. Answering them can require finding brief moments scattered across hours of footage. Agentic models that decide when and where to look promise to find these moments, yet this ability is hard both to measure and to learn. Sparse supporting evidence may establish that an action occurred without establishing its total count; verifying timeline-wide claims requires checking for missed occurrences and double counting. Answer-only supervision, in turn, does not explicitly supervise the search for evidence. We present NexSight, which builds verifiable tasks to both evaluate and train agentic long-video understanding. Its agentic pipeline, SightForge, accepts a task only if a verifier blind to the draft answer can reconstruct it from evidence densely reacquired from the raw video and counterexample checks fail to refute it. This yields SightLine, 310 evidence-traceable tasks from 108 long videos, evaluated through SightLoop, a shared multiscale harness supporting temporal exploration, local inspection, evidence integration, and revisiting while recording returned observations. On disjoint videos, solver trajectories on verified tasks become step-level supervision for observation decisions. The strongest model reaches only 54.52% on SightLine, yet 79.35% when given the annotated evidence. With under 6K examples, Sight-35B nearly doubles its base model's interactive accuracy (21.29% to 42.26%), outperforms the 397B model of its family, begins using explicit spatial inspection after post-training, which its base model never uses, and gains over 8 points on two external long-video benchmarks.
This yields SightLine, 310 evidence-traceable tasks from 108 long videos, evaluated through SightLoop, a shared multiscale harness supporting temporal exploration, local inspection, evidence integration, and revisiting while recording returned observations.
This yields SightLine, 310 evidence-traceable tasks from 108 long videos, evaluated through SightLoop, a shared multiscale harness supporting temporal exploration, local inspection, evidence integration, and revisiting while recording returned observations.
The strongest model reaches only 54.52% on SightLine, yet 79.35% when given the annotated evidence.
With under 6K examples, Sight-35B nearly doubles its base model's interactive accuracy (21.29% to 42.26%), outperforms the 397B model of its family, begins using explicit spatial inspection after post-training, which its base model never uses, and gains over 8 points on two external long-video benchmarks.
2 tiếng nói · chưa gắn nhãn trưởng lab · 2 nền tảng · chưa có số tương tác · Nguồn crawl
Running a frontier-scale model locally means keeping all its weights in GPU memory; the largest models no longer fit on one machine. Two NVIDIA DGX Stations... Running a frontier-scale model locally means keeping all its weights in GPU memory; the largest models no longer fit on one machine. Two NVIDIA DGX Stations connected over ConnectX-8 RDMA serve one model through a single endpoint, with tensor parallelism splitting the model weights and computation across both GPUs. The NVIDIA GLM-5.3 NVFP4 model, a quantized version of Z.ai’s mixture-of… Source
2 tiếng nói · chưa gắn nhãn trưởng lab · 3 nền tảng · 4 tương tác đã đo · Nguồn crawl
16 tiếng nói · chưa gắn nhãn trưởng lab · 1 nền tảng · 12 tương tác đã đo · Nguồn crawl
Streaming video LLMs can continuously observe the world, yet often handle user requests in isolation. We introduce ALIVE, an interactive streaming video model supporting proactive monitoring, streaming task coordination, and feedback-seeking interaction. We construct ALIVE-SFT-288K and ALIVE-RL-3K for two-stage training. Contrastive supervised fine-tuning teaches when and how to respond, while reinforcement learning with Behavior-Aligned Credit Assignment (BACA) combines trajectory-level and local advantages to refine decisions, responses, and silence across multi-turn interactions. We also introduce FSI-Bench to evaluate feedback seeking and subsequent task completion. ALIVE achieves the highest overall scores among the compared models on four public benchmarks and FSI-Bench, demonstrating strong streaming understanding and interactive capabilities.
4 tiếng nói · chưa gắn nhãn trưởng lab · 1 nền tảng · 2 tương tác đã đo · Nguồn crawl
Visible artifacts are an unreliable cue for detecting generated videos, because newer generators learn to remove them. We instead detect a trace left by the generation pipeline itself. Most video generators decode latents with a 3D causal VAE that groups consecutive frames in time, so the fidelity of generated frames rises and falls periodically with their position in a group. We call this periodic trace VAE Rhythm. Real videos are captured frame by frame and do not have it. To reveal it without training, we propose a phase scan. The scan reconstructs a clip with an off-the-shelf VAE from successive start offsets and measures the periodicity of the reconstruction error across these offsets. We find VAE Rhythm in most of the 39 variants of 31 generators we test, including the closed-source Sora, Kling 3.0 and Seedance 2.0. We use the rhythm as a zero-shot score, as a plug-in for existing detectors and as a data engine that pairs real videos with their VAE reconstructions to train our detector CaDet. On DF26, the training-free rhythm score outperforms all 15 retrained baselines, and as a plug-in it improves all 19 tested detectors by 6.7% to 23.7% AUROC. CaDet achieves the best AUROC on AIGVDBench, DF26 and three 2026 closed-source models, and its variant trained without any generated video still exceeds 90% on AIGVDBench.