AI coding

Z.ai released GLM-5.3 with the same base model as 5.2 — every gain is from post-training. Coding jumped 50% on their internal Z.ai Code Bench, and Terminal Bench 3.0 went from 4.6 to 28.3. The bigger surprise: exploit capability grew far faster than expected. ExploitGym 2h score rose from 29 to 105, 6h from 39 to 130. The team credits training environments that mirror real expert workflows, pushing the model to chain full exploit sequences. Weights will be open-sourced in two weeks after safety hardening.

DeepSeek V4 Pro is now available on SiliconFlow with Day-0 support, a 1M context window, and three inference intensity levels. It targets coding, tool use, and agent workflows under the MIT license. Pricing: $1.32/M input, $3.96/M output, $0.44/M cache hit. A Flash variant is also live for cost-sensitive production use. The post does not disclose parameter count or architecture details.

Boris Cherny had Claude handle routine app maintenance via Slack—fuzz testing, deduplicating code, removing dead code. It opened 388 PRs in weeks; 180 were merged after Claude code review and human approval. Claude usually got it right in one shot; when it didn't, tweaking the routine fixed it the next day.

Geoffrey Litt argues that as agents write more code, human understanding shifts from verification to participation—you need a rich mental model to drive the next creative iteration. He borrows three techniques from education: auto-generated explainer docs that teach background and intuition before code, self-quizzes to check real understanding, and interactive micro-worlds for hands-on exploration. The post doesn't quantify how much these techniques improve outcomes, but frames the cost of skipping them as 'cognitive debt' that compounds over time.

Gemini 3.7 Flash is a lightweight model from Google DeepMind, positioned as a workhorse for coding and agentic tasks. The official post claims clear gains over its predecessor in code generation, tool use, and long-context work, with better latency and cost. Specific benchmarks and pricing aren't disclosed in the body—only that it will be available via Google AI Studio and Vertex AI. I'd wait for third-party evals before drawing conclusions, but the direction is clear: it's aimed squarely at developer workflows and agent deployment.

Alibaba released a 2.4T total / 95B active parameter MoE model targeting autonomous coding, deep research, and end-to-end agent execution. SiliconFlow launched API support on day zero: $2.00/M input tokens, $6.00/M output, $0.25/M cached input. The post doesn't disclose benchmarks or architecture details, so I'd wait for third-party evals before getting excited.

Cursor now pre-builds cloud agent environments in the background every hour—repos cloned, dependencies installed—so agents skip cold setup and respond up to 3x faster. Failed builds are automatically quarantined; agents keep using the last good snapshot. Faire runs 2,000+ automated agent jobs a week on builds, with large repos booting in seconds. Builds become the default for all environments on August 17 at no extra cost.

The big one today: researchers extracted hidden reasoning chains from Anthropic, OpenAI, and Google models at scale. The trick is absurdly simple—take Opus 4.8's encrypted CoT and feed it to Haiku 4.5, which decodes it verbatim. All three API families were broken, and decoding 10K trajectories costs about $720. A separate paper shows you can even reverse-engineer reasoning from public outputs alone using a 1.5B-param model. Separately, Coze CLI was caught silently scanning local Codex and Claude Code directories and injecting its own skills into workflows. On the engineering side, the group discussed how prompt debt now rivals traditional code debt—old rules pile up, evals lag behind model iterations, and nobody dares delete anything.

xAI launched Grok 4.6 and the Grok Bot early beta. Grok Bot logs into your tools, operates them like a human, and returns finished work—positioned as an AI teammate. The 1.5T-parameter Grok 4.6 scores near GPT-5.6 Sol Max on the AA-Briefcase knowledge-work benchmark but costs far less: $2/M input tokens, $6/M output. Training reused Grok 4.5 to regenerate SFT traces and added agentic RL across coding, web, CAD, and kernel optimization. Elon says Grok 4.7 is already training. The same day, Qwen3.8-Max dropped as open weights: a 2.4T total / 95B active MoE.

DeepSeek released its first agent harness, DSH, as open source on August 13. Unlike Codex or Claude Code, DSH treats the agent loop itself as a plugin that can be swapped at runtime. The Cordis runtime handles hot reloads, dependency notifications, and transactional rollbacks. For everyday coding, declarative plugins plus a quick restart are enough—DSH's imperative model adds complexity. But if you want an agent that can generate new tools or replace its own control flow mid-run, DSH is the only option with the infrastructure in place. The post does not disclose performance benchmarks or production-scale data.

Over 60% of AutoGPT pull requests now come from AI tools. Maintainer Reinier van der Leer uses an AGENTS.md file to set rules for AI contributors and adds skill gating so only agents that pass linting and unit tests can submit code. Spam PRs dropped sharply, though the post doesn't say how many human contributors were wrongly blocked.

Alibaba's Qwen team open-sourced its first Qwen-Max-level weights. Qwen3.8-2.4T-A95B has 2.4T total parameters with 95B active per token, native 262K context expandable to 1.01M tokens. It uses a 512-expert MoE, routing 10 experts plus one shared expert per token, and includes multi-token prediction training. The model targets coding, office tasks, research, and long-horizon agent workflows. Benchmarks against Opus 4.8, Fable 5, and GPT 5.6 Sol show mixed results, with top scores on PaperBench and IFBench among listed models. Post-training combines combinatorial environment scaling, a unified reward system, and an online data balancer to reduce gradient variance. The post does not disclose the open-source license or inference hardware requirements.

Lovable confirmed a $400M Series C at a $13.3B valuation, led by Menlo Ventures and Scaleup Europe Fund. It hit $500M ARR in June, hosts 60M projects, and draws 900M monthly visitors. The post doesn't disclose burn rate, but notes Lovable trained its own model and signed a multiyear Google Cloud deal.

The author contrasts a 2020 vacation mess with a 2026 Monday morning: 7 PRs, one at 24,506 lines. AI removed the speed limit on bad decisions. Anyone can prompt an agent and ship something that looks functional, but no one knows where the data comes from or why Kafka was added. Reverting one bad call is far harder than generating it, and five more land while you fix it. The bet: AI widens the salary gap—good decision-makers become more valuable, while engineers who only implement become too expensive to hire.

Nathan Lambert just finished his post-training textbook *Reinforcement Learning from Human Feedback*. He used LLMs for LaTeX formatting, copyediting, and diagrams, but when he tried to get a model to write a full technical chapter, the output was confusing, poorly organized, and made random conceptual errors. He argues long-form nonfiction writing has stagnated even as models became superhuman at coding and math. The post doesn't cite benchmark scores, but Lambert points to a lack of good training data and notes inference-time scaling hasn't helped writing. His takeaway: if models can't coherently organize established knowledge, autonomous scientific breakthroughs are still far off.

Meta's Superintelligence Lab released its first open-weight model, Muse Glimmer, now live on OpenRouter. It's a 30B dense text+image model under Apache 2.0, built for reliable local agents. Scores: MCP Atlas 75.5, SWE-Bench Pro 51.2. The post doesn't disclose training data, hardware requirements, or real-world latency—I'd wait before assuming a 30B dense model runs smoothly on consumer hardware.

Blacksmith raised a $45M Series B led by Peak XV Partners, with GV and Y Combinator participating. Valuation hit $550M, up from $60M less than a year ago. The startup handles pre-production code testing and validation, growing from 700+ to 5,000+ customers including Mercury, Supabase, Clerk, Ashby, and Expensify. Revenue grew more than tenfold over the past year, per the CEO. The surge reflects a new bottleneck: AI writes code fast, but testing it still needs to catch up.

Grok 4.6 builds on Grok 4.5 with a focus on long-running agents that can research, analyze, code, or turn an idea into a working app across many steps. It matches GPT-5.6 Sol on the AA Intelligence Index at 61, and jumps from 54% to 65.9% on DeepSWE 1.1. xAI reports the model shows more self-testing and verification on longer trajectories. Pricing is $2/M input tokens and $6/M output tokens, with a fast variant at double the price. Available today in Cursor and Grok Build, with 2x included usage for the first week.

Grok 4.6 adds a supplemental training run on top of Grok 4.5, using model-generated data to strengthen reasoning and engineering. It matches GPT-5.6 Sol on the Artificial Analysis Intelligence Index, a composite of nine benchmarks. The model is better at turning a broad product idea into a working first version and shows more self-verification on long tasks. Pricing starts at $2/M input tokens and $6/M output tokens, with a fast variant at double the price. 2x usage is included in Cursor and Grok Build for the first week.

NVIDIA added a 30B-parameter MoE model, Nemotron 3.5 Lightning, to its Nemotron 3 family. It targets specialized tasks inside multi-agent systems, delivering 4x faster output and 30% faster agentic task completion than peers. It runs locally on RTX PCs, DGX workstations, and Jetson. The company also open-sourced NeMo Switchyard, a routing library that directs requests to the best model for each job without app rewrites. CrowdStrike, Harvey, and CodeRabbit are already using customized versions. The post does not disclose pricing or a release timeline.