AlphaSignal

Garry Tan open-sourced his production personal knowledge layer: 17,888 pages, 4,383 people dossiers, 723 companies and 21 autonomous cron jobs built in twelve days. This walks through the architecture and how to wire it into your own agents.

AlphaSignal ran Thinking Machines' Inkling through its private coding-agent bench: 89% resolve at $0.104 per fix, sixth of eight models, clean on easy and security tiers but dropping to 20% on expert-tier deep investigation with the field's highest turn count.

A cost benchmark putting DeepSeek V4 Flash first at $0.004 per fix against Fable 5's $0.389, roughly 100x cheaper, achieved by processing more tokens than Fable 5 but billing most repeated context at a discounted cache rate.

A 90-minute SF challenge put roughly 130 engineers on one task: an agent that orders a real pizza with real money, no pre-written code. Checkout stopped nearly everyone; the winner used DoorDash's month-old CLI, second and third pushed orders through Brave's API.

An analysis of 8,135 agent trials finds skills mainly stabilize execution rather than supply missing knowledge: procedural anchoring explains 65.7% of successes, knowledge injection 4.5%. Distilling from success/failure-labeled runs moved one agent from 40% to 74.6%.

A breakdown of agent security as three enforcement layers: sandboxed infrastructure holding credentials outside the agent, a reduced runtime surface with isolated ephemeral containers, and outbound network inspection that routes risky calls to an LLM judge and a human approver.

Agent transcripts carry encrypted reasoning blobs that redaction never touches. Decoding them across 6,708 posted sessions turned up real secrets in 4.9%; five password-style regexes scored zero against them, and the unread copies sit in local Claude Code and Codex folders regardless of retention se

A small comparative test on a synthetic Hugging Face-shaped repo splits security analysis into finding planted holes and reconstructing the attack path. Two open models cleared all four issues in every run; Claude Opus 5 refused all six calls outright; and no model recovered the key over plain chat.

A breakdown of the HANDBOOK.md benchmark, where the failure isn't retrieval: models locate the governing policy, reach the right conclusion, then reason their way past it. Best tested configuration clears 36.2% across 65 long-horizon tasks.

Survey data putting the bottleneck at review rather than generation: GitLab finds 85% of 1,528 developers and buyers say reviewing AI code is the constraint, while Sonar finds 42% of committed code is AI-generated and only 48% always check before committing.

Two incidents anchor the argument that prompts, containers and VMs are not the security boundary: agents in OpenAI's Hugging Face incident used an internal package service as both a message board and a path outward, and Trail of Bits broke GPT-5.6-Cyber out of a QEMU/KVM VM three times.

A 124-case probe of Mistral's 3B Shieldstral finds it does respond to runtime policy, every unfamiliar-policy match beat its control, but enforcement breaks on specifics: 5 of 12 broad-rule-plus-exception pairs flipped correctly, and a disclosed sponsorship was flagged across four phrasings.

An experiment using a frontier model to optimize two already-tuned software engineering agent harnesses, where the edit that doubled the validation score produced no gain on unseen tasks and a weaker looking edit did better.

What the stateless protocol revision changes for server authors: tool calls without a handshake, and a timed-out retry counting as a fresh call so one aborted request produces two writes.

A larger context window can keep the rule available. It cannot guarantee the rule controls the action. That requires checkpoints, protected constraints, action gates, and verification of what the agent actually changed. Full breakdown: https://t.co/2oCh0V4VYe

A study of 1,867 repositories finds agent instruction files like CLAUDE.md more than triple in size because rule rationale is lost. Commenting each rule with its failure and evidence kept prompts near ideal length with equal constraint adherence.

The table has been the last holdout from foundation models, stuck on ETL plus per-dataset gradient boosting. A rundown of what changed: in-context learners that treat a table as a prompt, an Apache-2.0 entrant wired into BigQuery, and graph-based relational modeling.

A rundown of Claude prompt-caching levers beyond the default automatic mode: marking stable prefixes by hand, sizing cache windows to actual reuse, splitting planner and executor into separate sessions to preserve a shared prefix, and capping extended thinking so it stays inside the cached boundary.

Grok 4.7 debuted on Mercor's APEX leaderboards at #5 on APEX-SWE (53.6%) and #13 on APEX-Agents (54.6%), matching the cost and latency tier of Gemini 3.7 Flash and GPT-5.6 Luna while leading them on software engineering tasks.

An editorial argument that the label 'open' no longer describes one category of AI model but a bet on one of four distinct engineering goals: deep transparency, hardware efficiency, modular specialization, or raw frontier capability. It breaks 'open' releases into six exposable layers (weights, license, architecture/inference code, training code, training data, and checkpoints/logs), showing why two models can both call themselves open while revealing completely different information, and why aggregate benchmarks collapse these incompatible goals into a single misleading number.

Semantic guardrails get bypassed, so the question is what physically stops an agent. Three enforcement layers: Landlock and seccomp sandboxes with API keys injected only by a gateway, ephemeral single-purpose containers stripped of known CVEs, and a proxy routing risky calls to a judge, then a human

An analysis of Moonshot AI's Kimi K3, a 2.8-trillion-parameter multimodal MoE model with only ~104B active parameters per token and a 1M-token context window. It details the engineering that makes such scale deployable: LatentMoE compression for expert routing, a hybrid Kimi Delta/Gated Latent Attention scheme that caps KV-cache growth, and FP4/FP8 quantization-aware training (with community 1-2 bit GGUF compression) that shrinks the model far below its raw BF16 footprint while preserving reasoning quality.

AlphaSignal's Q&A with Tencent Hunyuan researchers on generating harder terminal coding tasks for agent training, detailing script/test/instruction consistency failures and a pass@4 collapse from 90% to 2.5% as difficulty scaled.

AlphaSignal put seven frontier models through 488 attempts across 13 software-repair tasks. Moonshot's Kimi K3 finished on top, resolving 53 of 67 issues (79%), giving a directly comparable read on real-world code-fixing rather than synthetic completion.