DeepSeek V4.1 Flash-1017 Lands, and the Local Crowd Races to Run It

OpenCode users are reporting unauthorized API charges after a security breach, including one report of $400 spent on an Anthropic account; OpenCode confirms at least 1.4% of Console users were affected and its customer email warns about stored third-party provider keys. @sudoingX quotes the company calling it “a large scale distillation attack” and cites another user reporting $1,700. @MiaAI_lab, whose own account was drained of usage, urges affected users to change their API keys, and @sudoingX says he is now building a single-binary Rust harness so his keys stay on his own machine.

Addy Osmani argues that giving an agent a way to verify its work dramatically improves output; with that process, Claude Code wrote 90%+ of its own code. Uncle Bob’s Agentic Discipline 7 says a single disciplined agent can now take on big tasks that used to need a swarm-forge harness, and @poteto is working on a new verification skill for pstack. Matt Pocock’s prompt of the day: a pre-commit hook that blocks agent-authored files over 5,000 tokens and explains how to split them.

@poteto says the Grok Bot team has “been cooking and shipping” and points to what’s new and improved. @MiaAI_lab says it “keeps getting better and better” and promises to share workflows. @sudoingX is putting it to work reverse-engineering Hermes Agent’s self-improving loop and OpenCode’s speed into one lean Rust harness called xpress agent.

Ethan Mollick recommends the paper on how a team of Gemini 4 agents beat Minecraft as an accessible example of spontaneous coordination among agents, and of how their approach differs from humans’. He also asked Fable 5.1 to implement Borges’s short stories in 3D (after telling it quoting Borges was okay) and published the playable results.

NVIDIA released a new FP4 quant of GLM 5.3 Flash made with quantization-aware distillation, which fits on 2x DGX Sparks or 2x RTX PRO 6000s and, per NVIDIA, keeps almost all of the BF16 quality, @MiaAI_lab reports. Single-Spark owners can now run Qwen3.8 Flash Next on TensorFold with the new quant: 256k context, five concurrent users, 46-55 tok/s single-stream and 132 tok/s across five streams, with a later update adding around 14% to prefill.

Ant Ling dropped weights for Ling 3.1 Flash: 560B total, 25B active, 1M context. On its own chart it beats GPT-5.6 Sol on 10 of 11 benchmarks and goes 5-5 with Opus 5, @sudoingX notes. The model would not fit on his 2x DGX Sparks, so he messaged @niklaslenz_ai, whom he had never talked to, and borrowed his 4x Spark setup to load a Q4 build.

@sudoingX points out that the RTX 3060 is still the #1 desktop GPU on Steam and 16GB is now the most common VRAM, then lists what runs at each tier, including a 27B at 50.8 tok/s with 131k context on 12GB. He questions this week’s all-MoE releases, says the 20-30B dense class “has gone quiet,” and asks open labs for a 12B dense model. He also flags NVIDIA’s Nemotron 3.5 Bolt, a 74B MoE designed for RTX Sparks.

NVIDIA AI highlights Perplexity open-sourcing SPACE, the agent sandbox platform behind Computer, which gives developers more control over where agents run code and what they can access. It also posted a five-minute walkthrough of scaling a model endpoint on Kubernetes with NVIDIA Dynamo behind an OpenAI-compatible endpoint, plus a thread on partners post-training Nemotron 3.5 Super VL.

Two recent studies pitted human CPAs against AI models in tests of accounting acumen, Accounting Today reports. Both agreed that computers were still no substitute for flesh-and-blood accountants, “at least for now.”

In its tech news roundup, Accounting Today reports that BizBuyData will pay accounting firms for data to train AI, Clark launched Material Intelligence AI that stays active throughout an audit, and Digits launched Write MCP.

Given plain-language instructions, Claude Mythos 5 designed quantum circuits that improve on the best published ones, sometimes by an order of magnitude or more, Anthropic says. More efficient circuits reduce the resources quantum algorithms need, which could bring applications in chemistry and materials science within reach sooner. @kimmonismus quotes the work and calls it “RSI within reach.”

Satya Nadella says new Microsoft research in Nature shows machine learning can improve forecasts two to six weeks out, a historically hard window. His broader lesson: the future may be less about ever-more-capable models than about intelligent verification systems that learn where models are systematically wrong and calibrate predictions against reality.

@MiaAI_lab says Codex refused to review a plan written by Opus 5.5 after OpenAI’s cybersecurity content filter flagged it, and posted the filter message as proof. Separately, @kimmonismus notes that Epoch AI’s Cyber ECI shows the gap between Chinese open-weight and US closed models widening, from about 6 points in early 2025 to 11 now (GLM-5.3 at 160 vs. 171 for the US closed frontier).

Ethan Mollick warns that governments and companies are increasingly flooded with agentic AI inquiries that overcome the frictions that used to gate access, and says the graphs show it has already started. The consequential choice, he writes, is whether to make services worse or better in response.

A Chinese draft would require government approval to export the weights, checkpoints or code of the very largest models (2T+ parameters, 10^26+ FLOPs, or top 10 globally), @kimmonismus reports. API and app access are untouched so far.