Full GLM-5.3 Fits Four DGX Sparks at 319 GB

@bcherny tells users to prompt Claude the way they would brief a coworker: state the goal, say how much effort to spend, and describe how to verify the result. He runs no custom CLAUDE.md and keeps Opus 5.5 at medium reasoning effort. He also used Opus 5.5 with Lean to formally verify the Claude Agent SDK, and “a couple short prompts” produced 16 PRs fixing bugs and race conditions.

@coreyganim reports AI and his human assistant finish 50% of his daily work before he gets out of bed. Every Todoist task carries four fields: Why, Done, Access and Who. Where Who reads Codex, the agent completes the task, from drafting follow-ups to scheduling a UPS pickup. For the rest it drafts a brief in the task comments with phone numbers and context. @max_bevza calls good task context “the API between humans and agents.”

@VictorTaelin published OptMem, a 426-token prompt plus a roughly 600-line Python script that gives any agent permanent memory. The store is an append-only log where nothing is deleted, context stays constant-sized, and details fade with age. A new command expands a compressed memory into two larger parts so the agent can walk back through old recollections. He argues a memory sees at most log2(T) compressions, so errors have few chances to compound.

@jmacftw’s SparkGLM runs the GLM 5.3 Flash NVIDIA NVFP4 weights on two DGX Sparks through a fork of @AtlasInference’s Atlas, balancing quality, speed and concurrency. Against the TensorFold engine he calls speed generally comparable, with an edge on staggered concurrency. @alexellisuk questioned the pace for TP2, so McCartney posted the full RigMark results with prefill and KV size.

@free_ai_guides shares seven Grok Bot prompts for a bot connected to your bank. The audit groups 90 days of transactions and names the three fastest-growing categories. The subscription pass flags charges unused for 60 days and hands back cancel links without cancelling anything. The tax folder files receipts and invoices by month and builds a totals spreadsheet to send to an accountant. @1andypatterson reports his CFO bot found over $4,000 in fraudulent charges.

@LimestoneHQ counts 10 legal traps in weekend AI builds, citing 170 Lovable-built apps leaking user data through open databases. COPPA fines reach $53,088 per violation, CAN-SPAM the same per email, and California wiretap claims run $5,000 each. The EU AI Act has required AI disclosure from chatbots since Aug 2, 2026, with fines up to €15M or 3% of turnover. Some 3,117 federal website-accessibility lawsuits were filed in 2025, up 27%. Amazon paid $2.5B over auto-renew under ROSCA.

@Polymarket reports the Association for Human Mathematics is urging all mathematicians to boycott OpenAI after its release of 700+ mathematical research files. @ryanllm is skeptical of the group, writing it “sounds like something that spun up 3 days ago.” @trevorlasn pushes back the other way: mathematicians are still needed to verify AI claims. @rand_longevity flips it, calling for a boycott of human mathematicians.

@levie says the next stage of AI needs inference volume growing by orders of magnitude. His list: personal agents, swarm defenses for enterprises, security review of all code, and background agents running 24/7. Agents also need computers, networking and file systems alongside tokens. @ivanburazin quotes @auchenberg’s calculation that Muse needs 65,000 CPUs and 75PB of DRAM to serve 100 million users, about $2.8B of infrastructure for one app. Elon Musk replied, “People have no idea.”