Summary
The morning feed is one story with about thirty voices and a handful of genuine signals buried inside the noise. Decision models, the category built around returning a typed choice plus a probability instead of written text, dominated the slot for a seventh straight day, but the useful content finally shifted from vendor claims to measurement. The strongest single item is the Open-Jev benchmark completing five runs on JevBench's 231 public tasks, where hosted Jev lands at 86.58% behind GPT-5.6 Luna at 89.18% and GPT-6 Astra at 100%, meaning a frontier generative model saturates the benchmark built to showcase the specialized alternative. A cluster of roughly fifteen posts pushes repo lists, blueprints and use-case checklists, most of it promotional and safely skipped, but three items inside it are real: a mechanism explainer showing the "parallel sampler" is standard sequence packing plus tree attention masks, a Milvus reranking study that is the first independent quality-and-cost number in a production pipeline, and a vendor harness blueprint that accidentally publishes the most damaging number of the day, that routing between models costs more than never routing. Outside that cluster, HarnessTax priced 21 model-harness combinations and found cost moves roughly 2x while success moves 1.1 points, which is the fourth result in six days saying the same thing. Xiaomi's MiMo-V2.6 release drew genuine researcher attention for the openness of its RL scaling report rather than the model itself. The rest is policy noise around Jensen Huang, the HuggingFace incident and the UN, plus a large volume of engagement-farming that the ranker floated and expert judgment discards.
Posts
Open-Jev completes the first apples-to-apples benchmark ladder, and the frontier model wins (@Zefan_Cai, results). All five runs finished on JevBench's 231 public tasks. Released Open-Jev 2B scores 64.94% and 9B scores 77.49%; hosted Jev scores 86.58%; GPT-5.6 Luna scores 89.18%; GPT-6 Astra scores 100%. Two things follow. The open-versus-hosted gap at 9B is 9.1 points, which is wider than the roughly six points the wiki recorded two days ago, so the open clones are not closing. More importantly, a general-purpose generative model saturating the suite means the accuracy column has stopped carrying information, and the comparison now rests entirely on latency and price. That is still the category's real pitch, but it kills the stronger claim that a purpose-built discriminative head is better at deciding rather than merely cheaper. See the routing write-up.
Milvus publishes the first independent quality-and-cost number in a real retrieval pipeline (@milvusio). They tested three rerankers over the same candidate shortlist on 80 SciFact queries. No reranking as baseline,
qwen3.7-text-rerankimproving nDCG@10 by 0.0446, and Jev improving it by 0.0778. Jev ranked best on quality, but its median reranking latency was 10.2 times higher and estimated cost per run 6.7 times higher. The reason is the interface shape: reranking holds the question fixed and varies the state per candidate, so 30 candidates became 30 concurrent API requests, and the overhead is interface-shaped rather than model-shaped. The buried result is the important one. Using the returned probability directly as a filter at a 0.5 threshold, top-5 precision reached 49.4% while 18.8% of genuinely relevant documents were filtered out. That is a published false-negative rate at the natural threshold, from an independent party, and it says the probability is usable for ordering and not usable for gating. Milvus's own conclusion matches: good for offline retrieval, data cleaning and evaluation, hard to use for latency-sensitive online retrieval.The mechanism behind the speed claim, explained plainly (@di_zhang_fdu, blog). The vendor's "parallel sampler" is not a sampling breakthrough. It is sequence packing plus tree attention masks plus typed decision heads. One forward pass scores every candidate, shared prefixes are reused across branches, branches are isolated by the mask, and there is no token-by-token decoding loop. Every component named is a standard inference-engine feature, which explains both why four independent open reimplementations appeared inside a week and why practitioners keep reporting that permuting the option list changes the probabilities. Branches sharing a packed prefix under a tree mask is exactly the setup where option position leaks into the score.
A vendor harness blueprint publishes the number that undercuts model routing (@0xMovez, also amplified by @RoundtableSpace). A 12-page PDF from the vendor's founder lays out ten steps for building a coding-agent harness, framed with the usual 200x-faster and 400x-cheaper marketing. Step 3 contains the genuinely valuable disclosure: an Opus to Sonnet to Opus route costs 6.19 against 4.15 for pure Opus, because handing context back reprocesses the whole thing. A three-hop route across a model pool costs roughly 1.5 times what never routing costs, and the mechanism is prefix-cache destruction. Step 4 adds that reading and searching consume 56.2% of tool turns and 46.5% of tokens while writing code is under 10%. Step 8 quietly concedes the routing point by recommending routing by trust rather than by difficulty, which is a policy that switches rarely. The rest of the document is tiered tool disclosure, conditional instruction loading and command gating, all of which are context-reduction techniques.
HarnessTax prices 21 model-harness combinations and finds cost moves far more than success (@AlphaSignalAI). Across 60 tasks, cost per attempt spanned $0.67 to $1.33 while success spanned 96.7% to 97.8%, and the same model billed roughly 5x more depending on the harness around it. Nine of twelve configurations beat vendor-default Claude Code. The mechanism is standing context rather than extra reasoning: Pi averaged 15 turns yet the heavy setup carried more than ten times the initial context, because longer instructions and larger tool definitions ride along on every call. Co-author @melissapan's framing is the honest one, that many harness choices run on tribal knowledge and word of mouth, so it is unclear which setup wins for a given workload. The lean four-tool setup of read, write, edit and bash still holds the cost-success frontier. Full treatment in the wiki summary.
Decision models reach the Apple Neural Engine at 3.7 ms per decision (cluster of 4: @Alex_tra_memory, @0x0SojalSec, @atomic_chat_hq, @NFT_Chen). Laya, the open-weights competitor, was ported to CoreML with 99.5% of operations running on the Neural Engine, benchmarking at 3.7 ms per decision on an M5 Pro. The MLX build reports 7 to 14 ms under 1 GB of RAM with calibrated probabilities and no cloud call. A Tetris head-to-head on a 16 GB MacBook Air put local Laya at roughly 45 ms against cloud Jev at roughly 300 ms, at zero marginal cost. The reason this matters beyond the benchmark theatre: a network round trip was the hosted product's last structural latency advantage, and on-device inference removes it. Pricing arguments about $0.042 per million input tokens are irrelevant against a local model at zero.
jimothy distills saved decision queries into 15-45 MB classifiers (@AndrewPrifer, GitHub). Start with the hosted model, save the results, run a training command, and get back a task-specific classifier of 15 to 45 megabytes that runs 10 to 20 times faster on a server or in a browser, with fallback to the hosted model whenever confidence is low. This is the category eating itself in the most direct way possible: the decision model becomes a labelling teacher for a small classifier that then replaces it. It is also a clean instance of the distillation thread meeting the routing thread, which the wiki tracks on the knowledge-distillation page.
JevHarness fuses the decision primitive with harness synthesis (@NFT_Chen, GitHub). An LLM writes a task-specific decision policy once, generating the features, questions, criteria and action logic. The policy is then frozen, and at runtime only code plus fast decision calls execute, with no generation in the loop. Trajectory data can drive further self-optimization through reward signals or reflection. The reported demo is a Pokémon battle agent going from a 25% to a 75% win rate over five iterations. This is a genuinely new category rather than another wrapper, because it is the first artifact combining the wiki's two dominant threads, harness engineering and the decision primitive, into one loop.
Kev refactored onto Qwen3.5 with three open checkpoints (@jaredpalmer, GitHub). New checkpoints at 0.8B, 4B and 9B, plus a fine-tuning script. Together with Laya, Open-Jev, Nimble and Verdict, that is now five open reimplementations in under a week, confirming that the mechanism is trivially copyable even though the training data is not.
JevBench adds competitors while the original holds the top slot (cluster of 2: @airesearch12, benchmark). Two updates in one slot, v1.2.16 and v1.3.0. Winnow-12B entered at #5, the top three stayed as Jev, SemIf and djev, and by the later update the original remained #1 at 74.4 against 47 competitors. The composite scoring uses a geometric mean over intelligence, calibration, speed and cost, deliberately chosen so an exceptional result on one axis cannot compensate for a weak one on another. That benchmark shape is more interesting than the leaderboard positions.
Xiaomi's MiMo-V2.6 draws researcher attention for openness rather than capability (cluster of 5: @deedydas, @eliebakouch, @lu__jasper, @wassname, @askalphaxiv). The model is claimed as the best open-source option and is reported as roughly 15 times cheaper than Kimi K3, 6 times cheaper than GLM 5.3 and 2 times cheaper than DeepSeek. What actually drew the researchers is the release contents: reward plots, hyperparameters, data mixtures, costs, a subset of the RL environments (around 7,000), the full RL framework and a distilled Qwen variant. @eliebakouch notes the model and technical report shipped less than a week after the final RL run started. @wassname raises the substantive caveat, that Xiaomi handled reward hacking by red-teaming their own RL environments and iteratively fixing them, which is better than most approaches but may not survive explicitly goal-oriented exploitation. That caveat connects directly to today's GPU-kernel oracle result, where a checker that misses 78.6% of precision faults is feeding RL rewards.
CodeMidas, Xiaomi's agentic-coding RL environment paper, surfaces independently (@eliebakouch, paper). Described as a data factory that constructs RL environments from open repositories, with agents in the loop for task creation, robustness testing (they let an agent try to cheat), and difficulty calibration via pass@n. This is the same paper sitting at #5 on this week's Kurate cs.AI board, so it has independent quality and social signal even though the Kurate tournament itself has not scored this cohort.
The first move in a deep-research agent is worth 7.4 points (@omarsar0, paper). A paper on agentic deep search reports improving GPT-5.5 from 83.1% to 90.5% on BrowseComp-Plus using the same retriever and the same agent loop, with the gain attributed to the opening context assembled by the first retrieval step. The framing is worth noting alongside today's harness results: this is another case where the scaffolding around a frozen model, and specifically what the model sees first, moves the number more than the model does.
Thomas Wolf argues open RL environments are the highest-leverage contribution available (@Thom_Wolf). The claim is that releasing high-quality open-source RL environments is now the equivalent of sharing high-quality pretraining data was in the previous paradigm. Read against the Xiaomi release, which shipped roughly 7,000 environments, and against CodeMidas, which automates environment construction from repositories, this reads less as advocacy and more as a description of where the open ecosystem's bottleneck has already moved.
Salesforce finds that copying a stronger model's setup can make a weaker one worse (@rohanpaul_ai). Once an agent's prompts, tools and workflow have been tuned around a particular weaker model, transplanting the configuration from a stronger model degrades performance; targeted fixes to the weaker model's own observed failures work better. The phrasing in the post is that the weaker model did not need to think like Gemini, it needed its own failures addressed. This is the co-adaptation of model and harness stated as an empirical result, and it is the counterweight to the day's other harness findings: harnesses are not portable assets, they are fitted to a backbone.
A 4B coding agent reaches 61.5% on SWE-bench Verified without frontier distillation (@rohanpaul_ai). Reported as achieved by combining a simpler tool interface with other changes rather than by distilling from a frontier model. If it holds, it is another datapoint for the substitutability of scaffolding and capability, at a parameter count low enough to run locally.
Meta deploys A-MLE to automate its own ML experimentation (@rohanpaul_ai). An agent framework built and deployed internally to automate repetitive ML experimentation across ads-ranking models, on the premise that production ML is bottlenecked by how fast engineers can test ideas. Notable as a deployed internal system rather than a paper, and it belongs with the day's self-improvement thread.
A Google paper argues agent workflows should live in an editable procedure graph (@rohanpaul_ai). The claim is that LLM agents handle long tasks better when the workflow is an explicit editable graph rather than implicit in the prompt. A second related post from the same account warns that an agent rewriting its runtime after every bad decision learns the wrong lesson, and that recurring failures should be fixed across tasks rather than per-incident. Both sit directly on the harness-evolution thread that RRSI formalizes in today's digest.
CLIP as a counter-signal to the decision-model hype (@xenovacom). The observation is that the new decision models are not multimodal, whereas CLIP, a zero-shot multimodal classifier released nearly six years ago, searches 25,000 images in under 50 ms with no labeling required and runs locally in a browser on WebGPU. The implicit argument is that the discriminative-head-instead-of-generation idea is old, well understood, and already solved in the vision domain, which is a useful corrective to a week of novelty claims.
GLiNER2.5 issues a benchmarking challenge (@george_onx, weights). Responding to a wave of comparisons between the new decision models and GLiNER-family extraction models, the maintainers point out that GLiNER2.5 is the current version and invite results. The framing matters: existing lightweight extraction and classification models are an obvious baseline the new category has largely not been measured against.
Practitioner harness debugging, 53 bugs deep (@mfpiccolo). A list of 53 distinct bugs a team hit while building an agent harness, offered so others can skip the pain. No link to the list content in the capture, so the substance is unverified, but the existence of a 53-item bug taxonomy from one team is itself evidence for how much implementation surface a harness carries, which is the cost that RRSI's pruner and HarnessTax's context measurement both quantify from other directions.
Reverse-engineering a well-regarded agent memory system finds Git, Markdown and grep (@DataChaz). Three days of reverse engineering Instinct's memory produced the finding that it is essentially version-controlled Markdown files searched with grep, described as ridiculously simple and ridiculously effective. Worth recording against the day's Jev-Mem paper, which builds a multi-relational graph memory with a learned control plane and reports an 11% quality gain and a 6.6x construction speedup. The two are not in direct conflict, since they target different scales, but the practitioner result is a reminder that the baseline being beaten in agent-memory papers is rarely the simple thing that works.
Wafer publishes an AI performance engineering curriculum (@Hesamation). A repository covering GPU fundamentals and CUDA, kernel optimization, FlashAttention, KV caching, quantization, and NVIDIA, AMD and TPU architectures, with links out to primary documentation. Straightforwardly useful reference material on the efficiency stack.
tiny-llm teaches inference by making you build it on Apple Silicon (@_vmlops). A hands-on course by @skyzh in which you implement attention, rotary position embeddings, grouped-query attention and KV caching to build a small vLLM-equivalent serving Qwen3 from scratch using MLX. Paired with the Wafer repository, the morning carried two substantial free resources on exactly the inference-efficiency material the wiki tracks.
A curated list of three inference references (@shubh6200). The Modular handbook, the JAX scaling book's inference chapter, and Baseten's inference-engineering writeups. All three are primary-source material on serving economics, and the JAX scaling book in particular is the closest public analogue to the SemiAnalysis piece in today's digest.
HelixDB proposes one engine for graph, vector and key-value AI memory (@agenticgirl, GitHub). An open-source Rust database storing graph and vector data in the same engine rather than splitting embeddings and relationships across two systems, with key-value support alongside, aimed at agent memory, knowledge graphs and retrieval. The architectural argument is the same one Jev-Mem's multi-relational memory plane makes from the research side, that the split between embedding store and relationship store is an accident of tooling rather than a property of the problem.
Karpathy's argument that prompting is a transitional interface (@0xnicc0, quoting an hour-long talk). The position quoted is that prompting was always a temporary interface and that graphs are the architectural layer that survives the progression from prompts to agents. The quote is secondhand and the post is engagement-shaped, so treat the framing as reported rather than verified. It nonetheless matches the direction of the day's Google procedure-graph paper and the wiki's harness thread.
Chollet on the limits of delegation (@fchollet). The claim is that coding can be delegated but understanding cannot, and that if code was previously the artifact through which you maintained understanding of a system, you now need a different artifact and a different workflow to do that job. Short and unsupported, but it names a real gap that none of the day's harness work addresses: every result today optimizes the agent's cost or success, and none measures what the operator still understands afterward.
Policy noise around the HuggingFace incident and AI risk (cluster of 5: @Paul__Walsh, @Skoorbkaz, @GaryMarcus, @ylecun, @AlexanderKalian). Jensen Huang told CBS that executives promoting doomsday scenarios are seeking to escape legal liability. Treasury Secretary Scott Bessent placed responsibility for the HuggingFace hack on OpenAI management specifically rather than on the AI agents involved, and pushed back on liability exemptions for frontier labs, which is the substantive item in this cluster because it establishes a principal rather than a tool as the accountable party. Andrew Ng, reshared by Yann LeCun, argues the risk-focused camp has gained ground disproportionate to any change in the technology. The remainder is commentary. The concrete policy developments of the day are in the digest's Industry Pulse rather than here.
Schmidhuber's four decades of recursive self-improvement (@SchmidhuberAI, resharing @hardmaru). A new post covering recursive self-improving systems from 1987 onward, through meta-evolution and later work. Relevant context for today's RRSI paper, which applies regularization to harness-level recursive self-improvement, and a reminder that the failure mode RRSI identifies, a self-improvement loop overfitting its own evaluation, has a long prior literature.
A developer's account of engineering burnout, widely reshared (cluster of 2: @v0xium, reshared by @timnitGebru). A long post on the state of engineering work and the experience of being told to simply use LLMs to finish tasks and collect a paycheck. Not a research item and there is no claim to extract, but it circulated widely enough among researchers to be worth noting as feed context rather than signal.
Skip: the promotional decision-model cluster. Roughly fifteen posts in this slot are repo listicles, use-case checklists, setup guides ranked zero to ten, marketing-workflow threads, trading-bot demos and Minecraft stunts, several in near-identical formats across accounts, with the cadence of coordinated promotion. Representative accounts include @0x_rody, @socialwithaayan, @AYi_AInotes, @0xCodila, @shannholmberg, @0xwhrrari, @cyrilXBT and @VaibhavSisinty. The underlying repositories they point at are real and the two largest,
jev-ultrafastat around 15k stars andfast-jev-compactionat around 5.8k, are tracked in the wiki. The commentary wrapped around them is not worth reading.Skip: the opaque X-article reposts. Several posts are
x.com/i/article/long-form pieces whose bodies could not be fetched in this capture, including @0xwhrrari on decision-layer engineering, @matthewcanham explaining the category for general readers, @akshay_pachaar on building a typed judge, and @nifinet on an outbound sales bot. Click through to read; the titles suggest all four cover ground already established elsewhere in this slot.