TensorFold Recipes Arrive for GLM 5.3 Flash and Qwen3.8 27B

@robinebers asked Claude Opus 5.5 at high effort to build an interactive canvas of his app’s architecture, zoomable and clickable, after two months of shipping features faster than he could document them. The result drew connected graphs of what each part does, with step-by-step walkthroughs of flows like signup, payment and file upload through R2 storage and Modal workers. Replying to a flowchart skeptic, he notes the canvas lights up every major action on click, with an explainer for each step.

@PawelHuryn burned slices of popular plans and priced the calls at public API rates to put numbers on subscription value. OpenAI’s Pro 100 at $100 a month buys 5x the Plus allowance and Pro 200 at $200 buys 10x, downgraded from 20x. Claude’s Max 20x at $200 delivers 20x the Pro allowance, not 10x, and Meta’s Muse Code High Usage at $15 is 5x, not 3x. He ranks SuperGrok the cheapest western subscription for agentic work.

@plotarmordev reduces local decode ceilings to one division: memory bandwidth over data read per word. An M5 Ultra at 1.2 TB/s moving 5.3 GB per word at 5-bit caps near 225 tok/s. A DGX Spark at 273 GB/s moving 7.0 GB at NVFP4 caps near 39 tok/s. Speculative decoding lifts the Mac from about 114 to about 178 tok/s, and he notes one Spark recipe reaching 49 tok/s, above its ceiling.

@svpino reports that fine-tuning Qwen3 4B on AWS produced a model beating both the out-of-the-box model and Claude Sonnet 4.6. He argues a tuned small model can be cheaper and faster to run, and links a workshop codebase covering supervised fine-tuning, direct preference optimization and reinforcement learning from verifiable rewards. @cocoloba_ reports the same effect in vision: a small model trained on squash clips finds shots at 0.92 F1. The best frontier route he tested got 0.87, at roughly 1,000x the cost.

@alex_prompter notes everyone is writing with Claude Opus 5.5 this week, but GPT-6 Astra holds #1 on EQ-Bench’s creative writing leaderboard while Opus 5.5 sits 7th. His seven prompts for human-sounding output include a voice match over five past posts, an opinion-first format under 150 words, and an interview that asks three questions before writing. Others strip robotic lines, hunt for sentences anyone could have written, cut the draft to half its length while keeping every fact, and run a reply test from his harshest follower.

@AISystems_hq lays out 12 Claude processes for a solo business, from chasing invoices to closing the month. One rule covers all of them: Claude drafts, sorts, flags and files, and never sends, posts or pays without approval. The guide cites Xero research that owners spend about 36% of the week on admin, 15 to 20 hours. It also cites 59% of small businesses carrying invoices over 30 days late, averaging $17,700, with each process running as a skill on a schedule.

@DavidDecosimo flags an NYT report on Anthropic's courtship of religious leaders. It describes the company inviting religious leaders to San Francisco, having them sign NDAs, and making the case that Claude has a soul and moral standing. He calls the effort deeply troubling. In the replies, @brianchau57 cites earlier reporting that AI safety donors paid religious NGOs $3.3M for statements on AI.