# Jie Tang提出后训练缩放定律新观点，强调参数计数需结合数据、计算和运行条件，并介绍GLM 5.3通过合成RL环境取得重大进展 - 来源：Latent Space - 发布时间：2026-08-20 13:17 - AIWatch 分数：65 - AIWatch 标记：当日精选 - AIWatch 链接：https://aiwatch.icu/events/evt_01m0etd246qxkffwah76e4mgak - 原文链接：https://www.latent.space/p/ainews-death-of-params-zai-ceo-jie ## 精选理由 常规快讯，保留列表 ## AI 摘要 Jie Tang提出后训练缩放定律新观点，强调参数计数需结合数据、计算和运行条件，并介绍GLM 5.3通过合成RL环境取得重大进展。 ## 正文 We’ve covered GLM 5.2 very excitedly before, and Prof Jie Tang’s belief that there will be an open weights Fable-class model by end of the year (spot check - with 134 days left, there are now two 2-3T models (Qwen 3.8 Max and Kimi K3) with estimates that Fable is 3-7T, and only 2 points higher on the AA index.) Prof Jie Tang is back on X to tell us that our shorthand for model sizes is no longer enough: “Parameter count is only meaningful alongside three others — how much data you have, where you intend to spend your compute, and who will run the model, under what conditions.” X avatar for @jietang jietang@jietang Thoughts About Scaling Law Scaling, but not only of parameters. Every model release now ends with the same question: how many parameters? It isn't a question that can be answered on its own. Parameter count is only meaningful alongside three others — how much data you have, 5:04 AM · Aug 19, 2026 · 1.02M Views 170 Replies · 661 Reposts · 4.83K Likes We have covered Chinchilla (and post-Chinchilla) scaling laws in past LS years, but, so we will skip the history lesson, but it is good to level-set on why Chinchilla’s assumptions were wrong in the Inference Inflection world (no fixed number, between 200-900 toks/param, citing Roberts et al on task dependence). In short: Memorization prefers more parameters. Reasoning prefers more post-training data and effective depth. GLM-5.3’s big jumps come solely from RL on long horizon environments: The environments now cover a much broader range of production workflows, with tasks designed around how engineering and research work is actually carried out in practice. Some represent several days of work for an experienced engineer. In an ML infrastructure task, for example, the model may be given the same working environment as an engineer, with access to compute clusters, storage systems, internal documentation, codebases, and experiment results. It must diagnose bottlenecks across the training stack, implement optimizations, run experiments, and deliver a measurable end-to-end speedup while preserving correctness. Training on environments at this level pushes the model toward taking ownership of substantial work end to end, rather than relying on users to decompose the problem and supervise each step. For those following the recursive self improvement story, their entire environment and judging and verifier process is synthetic all the way down: As agent capability improves, much of the difficulty in scaling post-training moves from the model to the environment. A useful task environment has to be executable, verifiable, and close to real professional work — and we need many of them, not a handful of hand-built ones. To scale this process, we built pipelines that synthesize environments end to end, and for a subset of tasks, the RL reward signal as well. Research agents collect task patterns from real work and turn them into runnable long-horizon environments with multi-step dependencies and hidden state; a judge agent then attempts each task to verify that it is actually solvable. Verifiers are synthesized without access to the reference solution, while solver trajectories are used to discover and close reward shortcuts. A verifier that passes oracle, no-op, and unsolved-state checks produces a binary reward reliable enough to train on directly. To put an end to parameter count obsesssion, Prof Jie identifies 5 knobs of scaling, including MoE sparsity with the new XA-YB notation. He notes that advanced skills (e.g., finding software vulnerabilities) are not retrieval/memorization problems. They require carrying long causal chains (20+ inference steps) without losing the thread. This ability does not live in total parameter count once a certain knowledge-holding threshold is reached. And it looks like there is much more to go. AI News for 8/18/2026-8/19/2026. We checked 12 subreddits, 544 Twitters and no further Discords. AINews’ website lets you search all past issues. As a reminder, AINews is now a section of Latent Space. You can opt in/out of email frequencies! Open-Weight Models, Compression, and Benchmark Movement Ornith-1.5 lands as a serious new open family: @ornith_ released Ornith-1.5 in 9B dense, 35B MoE, and 397B MoE variants under MIT, with quantized formats including FP8, GGUF, MLX, and NVFP4. The headline claim is end-to-end self-improvement: the model proposes tasks, generates scaffolds, and produces RL rollouts to create new training experiences. Reported evals are strong across agentic/coding workloads, including Terminal-Bench 2.1: 86.1, SWE-Bench Verified: 86, DeepSWE: 56, HLE: 44.6, and Tool Decathlon: 71.2. The release was quickly wired into serving stacks by vLLM and Ollama. Compression continues to get more aggressive without fully collapsing utility: @UnslothAI and @danielhanchen shipped new Qwen3.8-27B GGUFs using Dynamic V3, claiming roughly 10% higher accuracy at the same size and releasing 1-bit quants that still retain about 77% of BF16 accuracy while running on 8GB RAM. Their new Divergence-300 metric extends top-1% greedy accuracy across longer generations using unseen examples from Terminal Bench, DeepSWE, and related tasks. Agent and legal eval boards continue to reshuffle: @arena published a Pareto view of Agent Arena, where Claude Opus 5 (High) leads quality, but lower-cost models like Kimi K3, GLM 5.2, Grok 4.5, and GPT-5.6 Luna define much of the value frontier. Separately, @ValsAI reported Grok 4.6 at #3/49 on Legal Research Bench with 48.1%, 500k context, tool/image/file support, and relatively low pricing. For open models, @ValsAI also highlighted GLM 5.3 as #2 on Terminal Bench, #3 on Legal Bench, and #6 on Skills Bench among open weights. Agent Harnesses Become the New Competitive Layer DeepSeek Harness’s minimalism is deliberate, not incomplete: A detailed writeup amplified by @ZhihuFrontier and summarized by @TheTuringPost frames DeepSeek Harness (DSH) as an intentionally thin shell over a plugin architecture called Cordis. The key design choice is that everything is a plugin, including the agent loop itself. Early beta users reportedly shipped 100+ plugins and filed 400+ issues in under a week; examples range from a gomoku model testbed to a database agent that closes the SQL feedback loop by connecting the model to live query execution. The strongest takeaway is architectural: DSH is less “productized assistant” than open agent runtime, optimized for user-extensible tooling, swappable control loops, and business-rule injection. TrueFoundry open-sources TrueForge and makes the harness-cost argument explicit: @truefoundry, @omarsar0, and @kimmonismus all covered the launch of TrueForge, an MIT-licensed, self-hostable, vendor-neutral harness for production agents. The stack includes tool orchestration, context management, subagents, code sandboxes, human approvals, and traces, with both local and hosted deployment modes. The technical claim that resonated: on a 14-task enterprise benchmark, TrueForge matched Claude Managed Agents on Opus 4.8 while using about 30% fewer tokens, and routing to GLM-5.2 cut cost by around 75% while preserving accuracy. The broader industry theme—also echoed by @bradenjhancock and @dbreunig via @rseroter—is that the session/environment/memory/tools layer is becoming a major source of both differentiation and savings. Managed harnesses are also getting sharper observability and controls: @ClaudeDevs added memory support for self-hosted sandboxes, domain allow/block controls for web tools, and a redesigned multi-agent session viewer with minimap, grouped transcript, and cost-per-thread/session. OpenAI, meanwhile, continues pushing the opposite angle: give teams the harness primitives to embed into their own products. @OpenAIDevs highlighted the open-source Codex harness as the runtime beneath internal tools, ops dashboards, and custom apps, while @cursor_ai shipped cloud-agent UX improvements around persistent goals and long-lived sessions. Post-Training, Mid-Training, and RL Systems Work More evidence that scaling is shifting from parameters toward training recipe quality: @kimmonismus surfaced a notable claim from the zAI/GLM founder: progress is still scaling, but too much discourse has fixated on parameter count rather than data quality, inference compute, and post-training. The cited example is GLM-5.3, reportedly based on the same core base model/architecture as GLM-5.2, but improved substantially via about one month of extra RL. Microsoft’s Agent Lightning points at RL-through-the-harness as a practical recipe: @omarsar0 highlighted Agent Lightning v1.0, which connects arbitrary harnesses to RL through an endpoint proxy, handling issues like retokenization, sample merging, advantage calculation, normalization, and scheduler/backend coordination. With ~6K training examples and modest compute, it reportedly moves Qwen3.5-9B on SWE-Bench Verified from 41.8% to 56.4%. Mid-training is being treated more explicitly as an optimization surface: @cwolferesearch laid out the current practitioner view of CPT/midtraining: optimize data mixture, duration, stage ordering, sequence length, and even post-trainability rather than just “continue pretraining on better data.” The thread is useful precisely because it frames these as interacting knobs rather than independent tricks. RL infrastructure keeps improving underneath the research: @SergioPaniego resurfaced work showing on-policy distillation in TRL becoming 40x faster via generation buffers, batched teacher calls, and binary logprob encoding; @mikasenghaas announced adaptive concurrency in prl, dynamically adjusting in-flight rollouts over the course of an RL run. Benchmarks, Retrieval, and Infra Details That Matter in Production Qdrant’s filterable HNSW vs ACORN is a substantive retrieval systems update: @qdrant_engine argued that filtered ANN should be addressed in the index, not only at query time. Their filterable HNSW adds edges between points sharing indexed payload values, keeping filtered subgraphs connected. In their benchmark on a 1% filter over 1M vectors, they report 99.8% recall at 1.0ms versus 67.7% at 4.7ms for ACORN. They also note ACORN still helps for broad values and AND filters, especially atop a graph already optimized for filters. Sentence Transformers v6.0 reflects the practical move from single-vector to multi-vector retrieval: @tomaarsen summarized the distinction clearly: dense retrieval compresses each text into one vector, while multi-vector retrieval keeps token-level vectors and scores query tokens against document tokens before aggregating best matches. That matters because late-interaction retrieval is increasingly the default tradeoff for quality-sensitive search systems. Production agent latency often has little to do with the model itself: @dair_ai summarized a paper instrumenting ten agentic apps and finding that non-LLM components dominate latency in half of them, with sandbox memory peaking at 28GB/session, up to 32x latency variation across subsystems, and long idle state retention between steps. The optimizations are unsurprising but important: task-aware serving cuts latency 29–40%, state offloading reduces memory 4.6x, and tool-result caching removes 35.2% of redundant search calls. Linear and turbopuffer show vector infra creeping into non-search hot paths: @turbopuffer said Linear moved its delta sync read path from Postgres to turbopuffer, using attribute indexes for permission filters and reducing the largest syncs by about 8 seconds. Google, OpenAI, Anthropic, and the Productization Race Gemini 3.7 Flash had a strong day on both evals and product integration: @_philschmid and @NewsFromGoogle highlighted Gemini 3.7 Flash taking #1 on Artificial Analysis’s AA-AnalystAgent, with 60.0% pass^5, 70.5% pass@1, 77.5% pass@5, 1.32s/task, and $0.54 average cost across 80 spreadsheet/document-heavy quantitative tasks. Google also pushed it deeper into product surfaces: Gemini chat and Spark, Search-based interactive simulations built on the fly in AI Mode (example), and AI Studio GitHub sync for build workflows. OpenAI is leaning into low-cost deployment and privacy positioning: @Replit launched Free Mode powered by GPT-5.6 Luna, which @kimmonismus framed as a meaningful efficiency win: a model that would recently have been SOTA is now cheap enough to be given away broadly. On the enterprise side, @OpenAI introduced Private Safety Processing, aiming to preserve Zero Data Retention for frontier models while still detecting cross-interaction safety risks without human access to the underlying content. Anthropic continues to tighten the developer ergonomics loop: beyond the managed-agent updates above, @ClaudeDevs added a Concise output style to Claude Code, another sign that product teams are now tuning not just capability but response-shape as a first-class UX variable. Top tweets (by engagement) Ornith-1.5 release: @ornith_ unveiled an MIT-licensed open model family from 9B to 397B, with strong coding/agentic benchmark claims and broad quantization support. OpenAI privacy/safety infrastructure: @OpenAI announced Private Safety Processing while reaffirming Zero Data Retention for frontier models. Gemini student push and product bundling: @GeminiApp offered a year of Gemini plans to students globally while rolling out new study-oriented features. Claude Code UX update: @ClaudeDevs shipped Concise mode, a small but widely noticed improvement for day-to-day coding-agent interaction. OpenRouter acquisition: @patrickc confirmed OpenRouter is joining Stripe, a move many interpreted as validation that token routing/marketplaces are becoming core infrastructure rather than edge tooling. Introducing Qwen3.8-27B Dynamic v3 Unsloth GGUFs (Activity: 1428): The image is a technical announcement graphic for “Dynamic v3.0 Qwen3.8”, showing Unsloth’s new Qwen3.8-27B Dynamic v3 GGUF post-training quantizations and claiming >10% higher top-1% accuracy at the same GGUF size versus other providers. It includes a memory table suggesting the model can run from 1-bit quants on ~8GB RAM up to BF16, plus a chart comparing accuracy across quant sizes; the post links the GGUF release on Hugging Face, the Dynamic 3.0 docs/benchmarks, and the image itself. Unsloth emphasizes these are post-training quantization releases only—“we do NOT use QAT or QAD”—and says the imatrix calibration file is public for independent evaluation and fine-tuning experiments. Comments were mostly positive, but one technical request asked Unsloth to add the previous UD 2.0 quants to the graph so users can compare against what they already have locally. Another commenter asked for deeper diagnostics, specifically per-category and KV-cache quantization KLD numbers, referencing localbench-style reporting.Several commenters requested more detailed quantization evaluation for the new Qwen3.8-27B Dynamic v3 Unsloth GGUFs, especially a direct graph line comparing against the prior Qwen 3.8 27B UD 2.0 quants. Suggested metrics included KLD and/or top-1 agreement, which would help users judge whether the new dynamic quantization is materially better than the versions many already have stored locally.A commenter asked for per-category KLD and KV-cache quantization KLD reporting, referencing the style of breakdowns from localbench.substack.com. This would make the quant quality discussion more actionable by showing which benchmark/task categories or cache-quant settings degrade most under different GGUF quant formats.There was interest in the practical memory footprint of the quants: one user noted ~15 GB for Q4_K_M, while another inferred that IQ4_XS may now fit on 16 GB VRAM “without mtp.” The technical concern is whether these smaller formats maintain model quality closely enough to justify running a 27B-class model fully on common consumer GPUs.