#Tutorial/How-to

AIMiniMax says its M2 series could not output the name Ma Jiaqi, a failure it traced to post-training data that rarely included the token. Its tests found the input embedding stayed stable while the lm_head weights for low-frequency tokens drifted during SFT. A synthetic full-vocabulary repetition dataset restored generation for affected tokens and reduced Japanese-to-Russian confusion from 47% to 1%.

Why it matters: The post traces a community-noticed token failure through tokenizer, embedding, and lm_head tests, showing how post-training data coverage can cause low-frequency token drift.

AIFunAudioLLM has released Fun-ASR-Nano-2512-hf, a Hugging Face Transformers-compatible version of its end-to-end speech recognition model, which supports Chinese, English, and Japanese. The Chinese coverage includes 7 dialect groups and 26 regional accents, and a separate Fun-ASR-MLT-Nano-2512 checkpoint handles 31-language recognition. Developers can run the model natively in Transformers 5.17.0 without custom model code or trust_remote_code=True.

AITri Dao says that after a mathematical rewrite, all transformer operations can be expressed as a series of GEMMs with epilogues. Given a few optimized primitives, LLMs and novice humans can write near speed-of-light kernels for transformer ops. The related CODA work fuses memory-bound surrounding ops into the matmul epilogue, and LLMs can also write CODA kernels approaching speed-of-light.

AISebastian Raschka reviews recent open-weight LLM architecture changes aimed at reducing long-context memory and compute costs. He covers KV sharing and per-layer embeddings in Gemma 4, per-layer query-head budgeting in Laguna XS.2, Compressed Convolutional Attention in ZAYA1-8B, and mHC with CSA/HCA compressed attention in DeepSeek V4. The article reports that DeepSeek V4-Pro uses 27% of single-token inference FLOPs and 10% of the KV cache size of DeepSeek V3.2 at a 1M-token context.

AICognition reports that deploying its Devin agent on hardware-in-the-loop and software-in-the-loop workflows cut failure triage time and multiplied test generation at automotive customers. One team reclaimed 2K–4K engineering hours monthly across about 4,000 tickets, while RV Tech rose from 1–2 to 10–15 generated tests per day. Devin also helps convert bottlenecked HIL tests into SIL equivalents to catch failures earlier.

AIFactory has released an Automated QA skill that drives an app as a real user would, filling forms, typing into terminals, and calling endpoints, then posts a structured report with screenshots, terminal snapshots, and API traces as a single updating comment on each pull request. Teams can run it on every push or make it an optional CI check triggered by a PR label, comment command, or manual dispatch, and developers can run /qa locally in any Droid session. Automated QA is available today in all Factory plans.

AITri Dao points to Wentao Guo's explanation of a mathematical rewrite of the MoE backward pass, which reduces activation memory and speeds up training, especially for fine-grained MoE. He also notes the work leverages new Blackwell features, including 2CTA MMA and CLC, to build fast MoE kernels.

AICognition reports that multi-agent setups work best when writes stay single-threaded and extra agents contribute intelligence instead of actions. It describes a code-review loop where a clean-context review agent catches bugs in Devin-written PRs, averaging 2 bugs per PR with roughly 58% severe. The post also says the smart-friend pattern, pairing a smaller primary model with a stronger one, has not yet worked well with asymmetrically weaker primaries and is an open training problem.

Why it matters: The post gives concrete findings on which multi-agent setups work, including clean-context code review and smart-friend escalation, and where they still fail.

AIA developer trained a 12M-parameter LLM on a custom ML framework built with a Rust backend and CUDA kernels, including Flash Attention, fused LayerNorm, and fused GELU. The framework claims 3x throughput gains, WebGPU fallback for non-NVIDIA devices, and a TypeScript API installable via npm.

AIZ.ai's GLM-5.3 delivers a 50% coding gain over GLM-5.2 on Z.ai Code Bench, reaching open-source state-of-the-art on public benchmarks including Terminal Bench 3.0. GLM-5.3-Flash uses 320B total parameters with 18B activated, combining linear and sparse attention to reduce compute and KV-cache needs. GLM-5.2 supports a 1M lossless context window for long-horizon tasks.

AIKarpathy describes using LLMs to compile raw source documents into a markdown wiki that he views in Obsidian, with the LLM writing and maintaining most of the wiki. He reports that at about 100 articles and 400K words, the LLM agent can answer complex questions directly from the wiki, and he also runs LLM health checks to find inconsistencies and gaps. He shares the underlying idea as an "idea file" that users can give to their own agents to build a customized version.

AIAndrej Karpathy describes using LLMs to compile raw research sources into a markdown wiki of about 100 articles and 400K words, viewed in Obsidian. He says an LLM agent answers complex questions against the wiki without RAG, with outputs filed back to enhance it. He also suggests the workflow could become a product rather than a collection of scripts.

AIAnthropic describes Claude Code auto mode, which delegates approval of agent actions to model-based classifiers instead of manual prompts or skipped permissions. The classifier reviews tool calls before execution and a separate probe screens tool outputs for prompt injection. Anthropic reports a 0.4% false positive rate on real internal traffic and a 17% false negative rate on real overeager actions.

Why it matters: The post explains the layered classifier design and its measured tradeoffs, showing how autonomous coding agents can cut approval fatigue without fully removing risk.

AIAnthropic's Labs team describes a three-agent harness with planner, generator, and evaluator agents for building full-stack applications over multi-hour autonomous coding sessions. The evaluator uses Playwright to test the running app against sprint contracts, and a retro game maker built with the harness worked end to end where a single-agent run's core feature did not. The author later removed the sprint construct and kept only the components still needed on Opus 4.6.

Why it matters: The post shows how a generator-evaluator loop, with explicit grading criteria and a tuned QA agent, turned a solo run's broken output into a working app, and how the harness was pruned as models improved.

AINano Banana 2, also called Gemini 3.1 Flash Image, adds visual grounding with Google Search, 512px resolutions, and extreme aspect ratios of 1:8 and 1:4. The guide advises using it as the default for new projects, with Nano Banana Pro reserved for complex prompts it fails, and keeping Thinking mode off by default.

AIHamel Husain and Shreya Shankar published evals skills, a set of skills for AI product evals that helps users avoid common mistakes. The entry point evals-start routes users to eval-audit for existing pipelines or error-discovery for unanalyzed traces. The repository is available at ai-evals-course/evals-skills and installs via npx skills add.

AIPublic AI benchmarks such as MMLU, SWE-bench Verified, and GPQA Diamond are saturating or showing contamination, prompting OpenAI to call SWE-bench Verified "no longer suitable" in late February and recommend SWE-bench Pro. OpenAI's audit found 59.4% of the problems its best model failed had flawed test cases, and GPT-5.2, Claude Opus 4.5, and Gemini 3 Flash could reproduce original fixes from memory. The article argues that behavioral tests, such as Vending-Bench's simulated vending machine business, may be more useful for everyday model choice.

AIMax Woolf published a blog post summarizing what he learned over several months working with agents including Opus 4.5 and beyond. The post also describes an optimization trick he considers promising.

AICognition reports merging 659 Devin PRs into its own codebase last week, up from 154 in its best week in 2025. The post describes internal workflows across web, Slack, Linear, CLI, and API, including Devin Review for PR diffs and bug catching, a daily design system audit, automated bug triage on Linear, and DANA for data analysis.

Why it matters: The post shows concrete workflows for using Devin across Slack, Linear, and code review, with specific usage figures that help teams judge fit for their own engineering processes.

AIThe article argues that engineers are splitting their work between building a harness of constraints, tools, and documentation for agents and directing the agents' work. It cites OpenAI, Stripe, and Anthropic examples, including architecture guardrails, custom linter messages, AGENTS.md updates, and plan-first execution. The author notes that open problems remain around code maintainability, verification at scale, and adopting these practices in older codebases.

AINicholas Carlini of Anthropic's Safeguards team describes an agent-team setup where 16 Claude instances worked in parallel on a shared codebase without human intervention to write a Rust-based C compiler. Over nearly 2,000 Claude Code sessions costing about $20,000 in API fees, the team produced a 100,000-line compiler that can build Linux 6.9 on x86, ARM, and RISC-V. The post focuses on harness design, including high-quality tests, lock files for task claiming, GCC as a reference oracle for the kernel, and the limits the project reached.

Why it matters: The post shows concrete harness design choices for long-running agent teams, including test design, locking, and parallel work division, that readers can adapt to their own autonomous projects.

AIAi2's Open Coding Agents family, with SERA as its first release, was built by Tim Dettmers and collaborators on 32 GPUs. The method generates synthetic bug trajectories with soft verification, comparing patches by line overlap instead of running tests. The post reports that a 32B model fine-tuned on about 7,000 trajectories for one private repository matched its GLM 4.5-Air teacher, and that the baseline costs $500 to run.

AIAnthropic's performance engineering lead Tristan Hume describes how a take-home test for hiring performance engineers was repeatedly defeated by successive Claude models. Claude Opus 4 outperformed most human applicants within the 4-hour limit, and Claude Opus 4.5 matched the best candidates in 2 hours. Anthropic is releasing the original take-home as an open challenge, with the best known Claude result at 1487 cycles.

Why it matters: The post traces how each Claude model defeated the take-home test, showing concrete redesign tradeoffs for evaluating engineers when AI assistance is available.

AITim Dettmers, a professor who has used Claude Code for eight months to automate his own work, argues that more than 90% of code and text should be written by agents. He says the coding-focused hype on Twitter overstates parallel sessions and autonomy, which translate poorly to most non-software tasks. The post offers a balanced guide to what actually works in agent-based automation.

AIAndrej Karpathy built hn-time-capsule, a tool that feeds each December 2015 Hacker News front-page article and its comment thread to GPT-5.1 Thinking for a retrospective analysis. The project, written with Claude Opus 4.5 in about three hours, processes 930 articles at a cost of about $58 and roughly one hour. Results include prescience and wrongness grades for commenters, and the project is hosted on his website with the intermediate data available for download.

AIThinking Machines is expanding free access to Tinker for researchers and instructors through a grants program. The program offers credits for scholars and students to fine-tune and experiment with open-weight LLMs.

AICognition says its autonomous coding agent Devin can complete a .NET Framework to .NET Core migration in as little as two weeks, using a Strangler Fig approach adapted from Jimmy Bogard's guide. The post says Devin handles planning via Ask Devin and DeepWiki, dependency sharing, controller and view conversion, and session state adaptation through a remote app.

AIThinking Machines published a post on on-policy distillation, a training approach combining the error-correcting relevance of RL with the reward density of SFT. The quoted post reports that in math reasoning and an internal chat assistant, on-policy distillation can outperform other approaches at a fraction of the cost.

AIThinking Machines Lab describes on-policy distillation, which samples rollouts from a student model and has a teacher grade each token with reverse KL. The authors report that this matches Qwen3-style reasoning results at a fraction of RL's cost, with AIME'24 reaching 70% in about 150 steps from a 400k SFT checkpoint. The method also helps recover instruction-following behavior lost during fine-tuning on internal documents.

Why it matters: The post explains why on-policy distillation gives dense per-token feedback, letting a small model match RL results at much lower compute cost.

AICognition rebuilt its Devin coding agent for Claude Sonnet 4.5, reporting 2x faster performance and 12% better results on its Junior Developer Evals, now available in Agent Preview. The team found the model is aware of its context window, which led to premature wrap-up behavior that they countered with repeated prompts and a 200k usage cap within a 1M token beta.

Why it matters: The post explains which agent behaviors changed under Sonnet 4.5, such as context-window awareness and note-taking, that forced a rebuild rather than a simple model swap.