Personal AI News Digest
32 topics from 2 sources
4,393 words · ~22 min read
A comprehensive framework tracking how machine intelligence production has shifted from human-made to model-made components since 2022. Stage 1 (2022): reward signals became synthetic via InstructGPT's preference-based training and Constitutional AI, with LLM-as-judge becoming standard evaluation. Stage 2 (2023): training data synthesis via Phi's textbook-quality generation and Apple's WRAP (3x pretraining efficiency), with NVIDIA Nemotron-4 340B shipping synthetic data pipelines and reasoning-trace corpora becoming standard. Stage 3 (2023): teacher models replaced human instruction via Alpaca ($600 fine-tune), Vicuna, Orca, and knowledge distillation, with DeepSeek-R1 making distilled model families the default. Stage 4 (2024): curriculum design became self-directed via Self-Instruct, STaR, Meta's Self-Rewarding Language Models, and SPIN, closing the loop so models generate tasks, judge outputs, and improve past human preference ceilings. Stage 5 (2026): research became automated via DeepMind's AlphaEvolve, Sakana's AI Scientist (Nature publication), and Karpathy's autoresearch—a minimal ratchet loop where a coding agent modifies LLM training, runs 5-minute experiments, keeps improvements only if validation loss improves, stacking 700 experiments into 20 kept improvements and cutting time-to-GPT-2 from 2.02 to 1.80 hours. Stage 6 (2026): RL environments became fully synthetic via Z.ai's end-to-end synthesis (research agents mine work patterns, judge agents verify solvability, verifiers stress-tested with oracle/no-op/unsolved-state checks) and Ornith-1.5's self-improvement (model proposes tasks and generates RL rollouts). Stage 7 (2025): human subjects replaced via Simile's post-training on interviews, transaction data, and registered RCTs to recover human bias and inconsistency, with Generative Agent Simulations achieving 85% accuracy vs. humans' 85% self-consistency, and SimGym/Tencent billion-persona approaches making focus groups and A/B tests inference workloads. Stage 8 (2026, in progress): physical world remains irreducible but compressible—Poolside's thesis that intelligence-bound problems are commoditizing while experiment-bound problems require real feedback; CZ Biohub imaging the Human Cell Atlas into virtual cells (1000x cheaper/faster than in vivo) with Chai, Xaira, and Lila's labs filling the AI-for-science stack. Across all stages, the pattern is consistent: synthetic versions are 10% worse, 100x cheaper, 10,000x faster, and improving on all three dimensions. Every flip was preceded by objections (model collapse, hallucination stacking, garbage-in-garbage-out) but happened when verification mechanisms made synthetic versions trustworthy—aggressive filtering for Phi, judge-vs-judge agreement for evals, unit tests for RLVR, oracle checks for verifiers, registered RCTs for simulations, wet-lab loops for virtual cells. The frontier advances not when generation improves but when verification does.
Sources AINews
Links [AINews] 10% worse, 100x cheaper, 10000x faster: Why Simulation is taking over, 2025 reading list, Z.ai GLM, the Poolside pivot, AI for Science themes, Simile pod, InstructGPT, Constitutional AI, Lee et al., LLM-as-judge, Textbooks Are All You Need, phi-1.5, WRAP, Nemotron-4 340B, Alpaca, Vicuna, Orca, on-policy generalized knowledge distillation, DeepSeek-R1, Self-Instruct, STaR, Self-Rewarding Language Models, SPIN, AlphaEvolve, AI Scientist, Nature, autoresearch, GLM-5.3 release, Ornith-1.5, Generative Agents, Generative Agent Simulations of 1,000 People, Open Science Framework, SimGym at Shopify, billion-persona, Chai, Xaira, Lila’s data-center-shaped labs, AI-for-science
The center of gravity in agent training is shifting from prompts to environments. GLM-5.3's sandbox-scaling interpretation shows the same base model achieves better long-horizon performance with richer executable environments and SAO-style counterfactual credit assignment. Google's EnvHarness/EnvRigger adapts static environments via plugin layers and policy-diagnosed reshaping, improving held-out performance by up to 9 points with 9.8% fewer execution steps. New benchmarks are task-specific and harder: FACET creates 6,078 executable terminal tasks from agent skills; SWE-bench Science introduces 119 scientific software tasks where Claude Code + Opus-5 is under 50% pass@1; CADBench finds top models at only 24.6% pass rate on realistic Fusion 360 tasks; AI4AI-Bench tests recursive self-improvement over 10 research repos with best model at 0.288 average score. On the inference-serving side, vLLM's IsoExec addresses a critical correctness issue: rollout/training logprob mismatches caused by floating-point non-associativity. It enforces bitwise parity across tensor parallelism (TP), expert parallelism (EP), and sequence parallelism (SP) layouts. On Qwen3.5-35B-A3B with DAPO on 8×H100, logprob difference dropped from 1.6e-2 to 6.7e-7 at 25.3% overhead—a necessary fix for reliable RL training where logprob consistency between rollout and training is load-bearing.
Sources AINews
Links GLM-5.3 release, Google’s EnvHarness / EnvRigger, FACET, SWE-bench Science, CADBench, AI4AI-Bench, vLLM’s IsoExec
Two inference-time architectural advances: DeepMind's Recirculation feeds contextualized deeper-layer activations back into earlier processing at inference time without retraining, reporting -60% contextualization errors, -23% perplexity, and +21% GSM8K improvement. Google DeepMind's Pandora's Router frames routing as an optimal search problem with costly inspection rather than assuming routing estimates are free, matching exhaustive-estimation quality while calling expensive estimators less often—applicable to specialist LLMs and variable inference-time reasoning.
Sources AINews
Qwen3.8-27B running locally on a single RTX 3090 with Unsloth Q4_K_S quantization, q8 KV cache, and 150k context demonstrates capable autonomous agent workflows: using Playwright with existing SSO/session cookies to navigate university systems and retrieve course schedules, and processing social-media video via download, frame extraction, Whisper transcription, and image enhancement. However, Qwen3.8-27B shows regression vs. Qwen3.6-27B on offline factual recall (Artificial Analysis' Omniscience benchmark), with failures on mildly obscure trivia and niche visual/history/geography tasks when web tools are disabled. This appears to be an intentional tradeoff: newer Qwen 3.x models optimize for coding, tool use, and agentic behavior rather than memorized factual storage. Commenters note the regression is acceptable given improved tool-calling and web-search workflows, with Gemma 4 suggested for trivia-heavy tasks. One user testing PI Agent vs. OpenCode on local Qwen3.8-27B reported PI Agent producing better outputs with fewer tokens, avoiding OpenCode's 32k output-token ceiling, and delaying context compression until ~90k tokens vs. OpenCode's ~67k at 100k total context.
Sources AINews
Links Qwen3.8-27b has the highest level of “agency” I’ve ever seen in a local model, image, Qwen3.8-27B took a serious hit to, Omniscience knowledge benchmark, Qwen 3.8 27b - PI AGENT vs OPENCODE, reddit.com/r/LocalLLaMA/comments/1j7r47l/..., Qwen3.8-27B took a serious hit to, Omniscience knowledge benchmark
DeepSeek-V4-Flash-Vision-Exp is live via the DeepSeek API (model='deepseek-v4-flash-vision-exp'), matching V4-Flash text capabilities while adding multimodal input. Benchmarks show broad gains over DeepSeek V4-Flash-0731: 83.9 on Terminal Bench 2.1, 75.9 on Toolathlon-Verified, 64.3 on Chartography, with Vision-Exp more competitive on multimodal tasks like Agents' Last Exam and ZeroBench; Opus-4.8 still leads many text-heavy benchmarks. DeepSWE reportedly improved by ~4 points from 0731 to Vision-Exp. The model supports Chat Completions, Messages, and Responses APIs with mixed text+image inputs via base64, external URLs, or the Files API; images are billed as up to 384 tokens each at V4-Flash pricing. DeepSeek also launched a Files API allowing users to upload an image once, reference it by file_id, and avoid resending payloads across requests, reducing bandwidth overhead. Model weights appear API-only at the time of discussion.
Sources AINews
Links DeepSeek-V4-Flash-Vision-Exp, mixed text+image API support with 117–384 image tokens billed at Flash pricing, Files API for reusable uploads, vision guide, files_api
A custom inference rig for DeepSeek V4 Flash-0731 uses 16× RTX 5060 Ti 16GB GPUs connected via two Broadcom/PLX PEX88096 PCIe switch islands with patched NVIDIA 610.43.02-p2p, Resizable BAR/BAR1 set to 16 GiB per GPU, and custom all-reduce/DSpark pipeline parallelism. Reported throughput is approximately 100–150 tok/s single-user generation depending on TP/PP layout, with concurrency scaling to 727 output tok/s aggregate for TP4/PP4 at 16 users. The setup demonstrates an unusual approach to distributed inference using consumer-grade GPUs and PCIe switching rather than traditional high-end interconnects, though the captured screenshot shows GPUs mostly idle despite high VRAM residency, suggesting the image is more a topology/memory residency proof than a live utilization benchmark.
Sources AINews
Links The boring way to run Deepseek V4 Flash-0731 130-150 tks - 16x5060ti 16GB over…, image
Deep analysis of how AI agent harnesses—the environment, tools, context, and guardrails surrounding model weights—have evolved and will continue to evolve. The piece argues that agents began working reliably around Christmas 2025 due to convergence of model capability improvements and harness maturation. The harness is defined as everything besides model weights that enables agents to perceive (context), act (tools), persist information (memory), and enforce boundaries (permissions).
Harness 1.0 ("Bolt-On Era") treated the harness as external scaffolding. Harness 2.0 ("Co-Training Era") shows measurable impact: Harness-Bench demonstrated a 23.8-point spread (52.4 to 76.2) on 106 tasks with zero model changes; OpenAI's GPT-5.6 Sol tripled ARC-AGI-3 score from 13.3% to 38.3% with only retained reasoning and compaction added. Models now absorb harness capabilities through RL training in real-world environments (e.g., codex-1). Anthropic recently deleted 80% of Claude Code's system prompt while retaining capability—the metric of harness evolution is how much can be shed as models absorb it.
Harness 3.0 ("Attention Era") predicts the harness will invert: as models absorb computer-facing capabilities (multi-agent orchestration, tool selection, memory), what remains are human-centric capabilities (permissions, identity, trust, legibility). The harness becomes the model's interface to human attention rather than the human's interface to the model. The author predicts within a year every agentic AI company will ship a "human attention policy surface" analogous to AGENTS.md, governing when agents interrupt, which decisions require approval, and learning from human corrections. The fundamental shift: tokens became abundant; human attention is the new bottleneck.
Sources Latent.Space
Links The Evolution of the Agent Harness, “Unsupervised Learning”, Harness-Bench, OpenAI achieved a similar result, recently deleted 80%, Extreme Harness Engineering for Token Billionaires, AGENTS.md
Simile AI, founded by Joon Sung Park (Smallville/Generative Agents author), raised $2B Series B backed by GreenOaks and Index Ventures with backers including Fei-Fei Li and Andrej Karpathy. The company runs tens of millions of simulations for Fortune 100 clients like CVS, achieving 85% accuracy in reproducing human behavior and attitudes compared to human focus groups. Simile's approach combines long-form interviews, observational and transaction data, randomized controlled trials, and post-training on causal mechanisms to build behavioral foundation models that capture how humans actually behave—not just what they say online. Unlike frontier LLMs (ChatGPT, Claude) which achieve 20–60% accuracy on niche populations by optimizing for rationality, Simile's models intentionally replicate human irrationality and biases. The company operates at scale of tens of thousands to hundreds of thousands of people today, with vision to eventually simulate all 8 billion people on Earth. Simile is observing early glimpses of scaling laws in simulation: more data about humans and more compute yield predictable gains in model performance. The company's vision extends beyond individual models to multi-agent simulations where agents interact in rich environments—similar to Smallville. For commerce use cases, agents operate independently; for society-scale simulations, agents communicate with each other. Long-term ambition is to simulate 8 billion people on Earth to answer emergent societal questions: climate change, democratic collapse signals, monetary system origins, and universal basic income (UBI). Simile estimates that in 5–10 years, simulations may cost as much as training a foundation model but could be so valuable (e.g., solving climate change) that the investment is justified.
Sources Latent.Space
Links Simile AI’s $2B Series B, Fortune 100 clients like CVS, 85–99% accuracy, Smallville, Simulation: the new Scaling L…, Simulation: the new Scaling Law — Joon Sung Park, Simile AI
Simile's core thesis is that frontier models lack complete mapping of 'social physics'—the underlying behavioral patterns of humanity. The company structures data collection into three buckets: (1) rich qualitative interview data capturing life stories, trauma, and personal context; (2) behavioral data from transactions and web scraping; (3) causal mechanism data from randomized controlled trials (RCTs) showing why people make decisions. RCTs are critical because most decision-makers want to know how to shape the future, not predict it—they need causal understanding. Simile partners with the Open Science Framework, which hosts tens of thousands of pre-registered, high-quality behavioral studies, to post-train models on RCT data. The company also runs its own experiments with real stakes (e.g., online stores where purchases are actually delivered) to ensure behavioral rather than merely attitudinal responses. Population-level and individual-level models are trained separately; both take subpopulation/individual descriptions and stimuli as input.
Sources Latent.Space
Links Simulation: the new Scaling L…, Simulation: the new Scaling Law — Joon Sung Park, Simile AI
Simile's validation approach is detailed in the 'Generative Agent Simulations of 1000 People' paper (published end of 2024). The study brought 1,000 US-representative participants to a virtual lab, collected two hours of wide-ranging data (interviews from American Voices Project, behavioral data), then created digital twins. After two weeks, participants returned for surveys and experiments (behavioral economics games, Big Five personality tests, General Social Survey, randomized controlled trials from PNAS). Digital twins predicted how source individuals would act in these studies with 85% accuracy—matching how well people replicate their own responses. Frontier models (ChatGPT, Claude) achieve only 20–60% accuracy on niche populations because they optimize for rationality; Simile's models intentionally replicate human irrationality and biases. A follow-up paper post-trained models on tens of thousands of RCTs from the Open Science Framework (a platform addressing replicability crisis in social sciences), showing significant improvements in behavioral prediction.
Sources Latent.Space
Links Simulation: the new Scaling L…, Simulation: the new Scaling Law — Joon Sung Park, Simile AI
Simile's product allows customers to query synthetic populations by filtering for specific demographics and environments (survey questions, behavioral experiments, A/B tests). Primary use cases include concept testing (evaluating different messaging, products, ideas), focus groups, and modeling earnings calls for public companies. The company has a strategic partnership with Gallup for policy research. Simile is intentionally cautious about political applications, wanting to establish guardrails before serving that market. Unlike simple polling (which tells you an outcome is bad but not how bad), simulation reveals the path to desired outcomes—potentially counterintuitive steps needed to achieve goals. Example: an automobile company marketing an EV might discover that the optimal strategy changes perception of non-EV cars, affecting overall sales in unexpected ways. Customers include CVS, Deloitte, and Wealthfront; Wealthfront was an early adopter of multimodal input (images, Figma mockups, websites) where agents traverse interfaces.
Sources Latent.Space
Links Simulation: the new Scaling L…, Simulation: the new Scaling Law — Joon Sung Park, Simile AI, Fortune 100 clients like CVS
Simulation differs fundamentally from prediction. Decision-makers rarely want to know 'what will happen'; they want to know 'how do we shape the future?' Prediction alone (e.g., 'sales will tank in two quarters') is unhelpful without causal understanding. Simulation reveals step-by-step paths to desired outcomes, sometimes suggesting counterintuitive moves. Park cites Asimov's Foundation series: psychohistory simulations showed that exiling scientists raising alarm about galactic collapse was the optimal first step—strange but necessary. In business, simulation might reveal that marketing an EV in a certain way changes non-EV perception and overall sales, an outcome invisible to prediction-only approaches. This is why causal mechanism data (RCTs) is critical: it shows not just correlations but the mechanisms by which interventions produce outcomes.
Sources Latent.Space
Links Simulation: the new Scaling L…, Simulation: the new Scaling Law — Joon Sung Park, Simile AI
Joon Sung Park's 2023 Smallville paper on Generative Agents (7,200+ citations on Google Scholar) demonstrated that AI characters could remember, plan, socialize, and develop emergent behaviors. The paper emerged from Stanford research in 2020–2021 when GPT-3 became available. Park and colleagues (Percy Liang, Michael Bernstein) played a 'time machine game' asking what single application would matter most in 10 years; they chose recreating the world through simulation over personal agents. The memory architecture used simple text/Markdown files rather than knowledge graphs or bespoke models, leveraging language models' strength at reasoning about text. However, limitations exist: retrieval and sense-making of large data is costly, and certain behaviors cannot be shaped by prompting alone—model parameters must be modified to learn new social physics. The paper's memory component was noted as underrated and became foundational for later work on behavioral foundation models. Park draws inspiration from Thomas Schelling's 1970s–80s agent-based modeling work, particularly the segregation model using red and blue dots on a grid. Schelling showed that small individual preferences for same-color neighbors led to complete societal segregation—a counterintuitive finding that informed mixed-income housing policy. Schelling won the Nobel Prize for laying groundwork for early simulations. Park sees an opportunity to revive agent-based modeling with high-fidelity generative agents: frontier models can now create rich descriptions of people, enabling complex societal simulations that could tackle wicked problems (climate change, democratic instability, UBI). If simulation successfully solves major societal questions, Park believes a Nobel Prize in Economics awaits.
Sources Latent.Space
Links Smallville, Simulation: the new Scaling L…, Simulation: the new Scaling Law — Joon Sung Park, Simile AI
Simile was founded by four co-founders: Joon Sung Park (researcher, Smallville author), Michael Bernstein (ImageNet co-author, human-centered AI pioneer), Percy Liang (coined 'foundation models' term), and Lainie Yallen (business lead from fast-growing AI-native companies). The company is ~60 people, headquartered in San Francisco (Mission Rock) with a newer office in New York. Nearly 20% of staff are Park's former lab mates from Microsoft Research who previously worked at OpenAI, Google Gemini, and similar organizations. Simile is positioned as both a research lab and product company: researchers continuously develop the vision while engineers, product managers, and designers build real-world applications deployed at Fortune 100 companies. The company is actively hiring across research, product engineering, and infrastructure roles, with preference for talent from admired companies (Figma, Notion, Rive) and personal networks.
Sources Latent.Space
Links Simulation: the new Scaling L…, Simulation: the new Scaling Law — Joon Sung Park, Simile AI
Park considered personal agents (automation tools that act on users' behalf) as a close second to simulation in the 'time machine game' exercise. However, he argues that accurate behavioral models must precede effective personal agents. A personal assistant that orders Hawaiian pizza for someone who dislikes pineapple fails due to lack of deep user understanding. Just as close friends have good mental models of each other, personal agents need rich understanding of users' preferences, values, and quirks. Current personal agents (OpenClaw, ChatGPT, Claude) leverage Markdown files for memory and are improving, but Park believes they lack the depth needed for true usefulness. He notes that some things cannot be shaped by prompting alone—model parameters must be modified to learn new social physics. While personal agents are advancing, Park contends that simulation (building accurate representations of people) is the necessary foundation.
Sources Latent.Space
Links Simulation: the new Scaling L…, Simulation: the new Scaling Law — Joon Sung Park, Simile AI
Poolside AI struck a non-exclusive licensing deal with NVIDIA for $6 billion plus a $1 billion investment at a $12 billion pre-money valuation. The structure is unusual: 109 Poolside employees are being hired by NVIDIA (receiving ~$6B in total compensation), founders stay with Poolside ($1B investment), and the Poolside Infrastructure Company (PIC) spinoff—building a 1.2 GW datacenter in Texas—continues separately. Founders framed this as neither acquisition nor acquihire, but a pivot driven by compute constraints. They noted losing a 40,000 GB300 cluster in January due to inability to close a $2B raise in a 6-week window, and that frontier-scale models now require "far more than an order of magnitude larger cluster" beyond 10,000–20,000 GB300s. The founders believe AI's ultimate value will shift from commoditized knowledge work (coding, accounting) to scientific discovery engines where real-world experimental feedback loops create data moats—a domain where "super intelligence will likely not" be commoditized by open source. PIC is positioned as a potential neocloud/compute provider, though founders said they are "not ready to share the updated vision."
Sources AINews
Links [AINews] Poolside gets $12B reverse-execuhire to NVIDIA; founders stay for $1B,…, Poolside’s Model Factory with Eiso Kant
OpenAI launched an Apple Messages plugin for ChatGPT Work/Codex on Mac, enabling message search, catch-up, drafting, and sending from the desktop app. On the builder side, ChatGPT Sites now support collaborative editing with teammates sharing projects while Codex manages git/CI, plus shared read-only conversation links and PR-context sharing to position ChatGPT/Codex as a coordination surface. Desktop memory/workflow features—Computer History, cross-app memory, and Record & Replay—rolled out to EEA, UK, and Switzerland for Pro/Business/Enterprise Mac users. The product strategy aims to capture user workflows on-device and turn repeated actions into reusable skills. On the API side, transparent backgrounds in GPT-Image-2 are now in preview for reusable design assets.
Sources AINews
Links @ChatGPT, @OpenAIDevs, shared read-only conversation links, PR-context sharing, transparent backgrounds in GPT-Image-2, @OpenAIDevs, Record & Replay
Anthropic announced general availability for computer use, browser tool, Skills API, and Files API on the Claude Platform. The Skills API adds versioned reusable procedures; the Files API now supports expiration control, 5x higher rate limits (500 RPM), and 1 TB/org. Anthropic also published an AG-UI adapter for Claude Managed Agents, mapping chat threads to managed sessions and streaming text, tool calls, and thinking into custom UIs.
Sources AINews
Links @ClaudeDevs, Skills API, Files API, AG-UI adapter for Claude Managed Agents
AT&T's internal AI deployment shows 40% of employee AI usage already routes to open models, with a target of 60–70%. Coding costs are down 56% for only a 2% quality drop, at 45B tokens/day. This supports the view that frontier closed models remain reserved for hardest tasks while "good-enough" open models eat the broad middle of enterprise demand, and signals a warning for OpenAI/Anthropic's enterprise moat.
Sources AINews
Links @Hesamation, @amir
GPT-5.6 Sol pricing dropped 50% via Router, and GitHub Copilot/VS Code users received temporary discounts. User sentiment indicates supply constraints are surfacing as usage caps rather than degraded quality: a $200/mo OpenAI Pro plan can be exhausted in a single heavy Codex day, and users report continuing to consume substantial tokens after hitting stated caps. On the open side, Ollama rolled out Kimi K3 to over half its subscription base with US/EU hosting and zero data retention. Gemma surpassed 1B downloads, with Google and community contributors (Awesome Gemma repo) aggregating variants, deployment guides, and fine-tuning recipes.
Sources AINews
Links @ollama, @eglyman, @github, @code, @bridgemindai, @theo, @ollama, @Google, @osanseviero, @_philschmid
Meta's Muse Spark 1.2 showed +2.1% net improvement in Agent Arena (up from 0.9% in v1.1), with particularly strong Bash Recovery (+11.4%). Third-party evals placed it #1 for Video-to-Website, #2 for Image-to-HTML, and #3 for Image-to-Frontend on DesignArena, sitting on the price-preference Pareto frontier. Meta previewed WildArtifactBench, an internal eval using win rates and Elo from human/agentic judges for practical multimodal tasks. Demos spanned visual coding, robotics planning, and audio-visual understanding.
Sources AINews
Links @AIatMeta, WildArtifactBench, @arena, @DesignArena
Zhipu's GLM-5.3 continues showing strength in agentic/code evals. GLM-5.3 Max shifts the Code Arena WebDev Pareto frontier, projecting to #2 among open models and #8 overall at 1597 points and $3.65/M tokens. It was integrated into AutoClaw, Z.ai's work agent. A key RL advance for GLM-5.2/5.3 is SAO (Single-Rollout Asynchronous Optimization), enabling stable asynchronous agentic RL.
Sources AINews
Links @AutoClawAIer, @arena, @ZixuanLi_
Gemini 3.7 Flash demonstrated strong cost-adjusted reasoning performance: ARC-AGI-2 at 84.6% for $0.25/task and ARC-AGI-1 at 95.5% for $0.12/task. It was also noted as excellent for agentic vision tasks.
Sources AINews
Links @arcprize, @JonathanJarvis
OpenAI's first NVIDIA Vera Rubin racks are now installed and running the training stack, explicitly tied to next-generation frontier pre-training. This marks a major milestone in the OpenAI-NVIDIA partnership.
Sources AINews
Links @udayruddarraju, @gdb
Cerebras' CS-4 essentially doubled performance on the same 5nm wafer, 4T transistors, and 900k AI cores via redesigned power delivery and cooling. Reported specs: 250 PFLOPs per WSE-3 Turbo, 43.2 PB/s memory bandwidth, and a 3-wafer CS-4 rack at 750 PFLOPs. Practitioners report 4,400+ tok/s per user on GPT-OSS-120B, up to 30x faster than GPU-based systems.
Sources AINews
Links @kimmonismus
Linux materially outperforms macOS for agent workloads, especially on filesystem-heavy operations. Semantic caching showed 57.1% hit rate, 55.7% fewer tokens, and ~15 ms hit latency. Gisting (context compression) achieved ~40% lower end-to-end latency and ~15% higher throughput with better results, per Shopify engineering writeup.
Sources AINews
Links @theo, @Qdrant_engine, @MParakhin, Shopify engineering writeup
Chroma launched Foundation, a research preview of self-improving memory built from prior agent sessions. This reflects a broader shift from "single-shot agent" thinking toward persistent harnesses with accumulated state, skills, and memories. Recent research highlights harness evolution (prompts, memories, skills, routing rules evolving independently of model weights) as more impactful than model weight updates. A key failure mode is harness-level forgetting: improving one component can silently break previously reliable behavior. Proposed solution: guarded harness evolution, separating proposing updates from committing them, with >10% gains across textual, multimodal, and open-world tasks. Negative results also matter: memory-based self-improving agents look worse once controlling for task order effects and evaluation variance; post-training agents tend to lock into an initial strategy early and spend remaining budget on local refinement rather than revisiting strategic choices.
Sources AINews
Links @jeffreyhuber, @omarsar0, @dair_ai, @omarsar0
Unsloth released Dynamic v3.0 GGUF quantizations of Qwen3.8-27B, claiming >10% better top-1 accuracy at the same model size versus other quant providers. Post-training quantization only—no QAT/QAD and no training on imatrix calibration dataset. Memory targets range from 1-bit quants runnable on ~8GB RAM up to BF16. Evaluation framed around Divergence-300 @32, KLD, and top-1% accuracy comparisons. Community requested comparative data against prior Qwen 3.8 27B UD 2.0 quants and per-category KLD metrics to understand where quantization loss appears across tasks and cache settings.
Sources AINews
Links Introducing Qwen3.8-27B Dynamic v3 Unsloth GGUFs, https://unsloth.ai/docs/basics/dynamic-3.0-ggufs, https://huggingface.co/unsloth/Qwen3.8-27B-GGUF
A home-built coding eval compared Qwen3.8-27B, DS4 0731, GPT-5.6-sol, Opus 5, Sonnet 5, and Haiku 4.5 across algorithm tests, repo bugfix/feature tasks, wall-clock completion time, and blind-judged code quality. GPT-5.6-sol leads overall with perfect repo-task performance and near-perfect algorithms. Local models are surprisingly competitive: Qwen3.8-27B scores strongly on hard algorithms and "surgical fixes" but is much slower; DS4 0731 achieves 8/8 on both repo tiers despite being a 2-bit local quantization. A practical tradeoff emerged: higher "thinking" improves some hard reasoning/code-quality cases but can overthink, increase latency, and even reduce repo-task accuracy compared with medium thinking. Community noted the benchmark may be saturated (all models near the top), making it hard to distinguish frontier/local capability, and requested more precise methodology and task definitions for reproducibility.
Sources AINews
Links I ran Qwen3.8-27B against Opus, Sonnet, GPT and others. Results inside.
Market research is a $100 billion industry, but Simile positions simulation as a tool for all human decision-making, not just market research. The company's TAM is ambiguous because it encompasses every decision made about humans for humans—a vastly larger scope. Park notes that as a researcher-founder, he didn't calculate TAM upfront; instead, he assumed that if simulation can inform every such decision, it must be valuable. The value case centers on replacing expensive human studies (e.g., Sam Altman's $40+ million UBI study in Africa) with instant simulations. Simile's pitch to enterprise customers emphasizes cost savings and decision quality: synthetic panels replace human panels at scale, ensuring stakeholder voices are represented in product and policy decisions. The company is impact-driven rather than valuation-focused; founders prioritize societal impact over financial metrics.
Sources Latent.Space
Links Simulation: the new Scaling L…, Simulation: the new Scaling Law — Joon Sung Park, Simile AI
Park draws a philosophical parallel between simulation and his background in oil painting (realism studios, figure painting). The best paintings teach something deep about their subject without being perfect representations; they highlight what matters most. Similarly, simulation aims to surface the essential nature of people and society. Park's artistic practice focused on mundane aspects of human life—a rural town study photographing people in everyday moments—reflecting a fractal worldview: understanding the inward (individual mundane details) reveals the outward (societal patterns). Simulation embodies this: modeling even the most ordinary aspects of people, when aggregated, teaches something profound about individuals and collective life. This perspective connects to documentary and biographical traditions (e.g., Studs Terkel's 'Working') that find compelling depth in ordinary lives. Park sees simulation as an exercise in understanding human society and collective existence, parallel to how AGI helps us think critically about human intelligence.
Sources Latent.Space
Links Simulation: the new Scaling L…, Simulation: the new Scaling Law — Joon Sung Park, Simile AI
Simile processes significant query volumes for large enterprise customers, making efficiency a consistent consideration. The company does not over-optimize early but thinks carefully about inference efficiency given the scale of multi-agent simulations and population-level queries. While sparse models and other architectural optimizations are not yet the highest priority, they are on the roadmap as deployment scales. The technology is already deployed at some of the world's largest companies, requiring robust infrastructure and cost management.
Sources Latent.Space
Links Simulation: the new Scaling L…, Simulation: the new Scaling Law — Joon Sung Park, Simile AI