Grok 4.5, Fable ADRs, GPT-Image Pipelines — AI Daily Jul 08

700 messages · 87 active members

700
messages
87
active members
@tounano, @thewildzeno, @samb69
top contributors

Overview

Two big model moves shaped the day: Grok 4.5 dropped with a meaningful jump in tool-chain planning, 500k context (1M coming), and cleaner agent behavior — Elon claims Opus 4.7-level, and @jcartu and @geilt confirm it feels transformational. Simultaneously, Anthropic extended Fable 5 through July 12 on Claude plans (up to 50% of weekly limit), widely read as a hedge against Thursday's GPT-5.6 Sol launch. Builders are burning tokens on skills-writing, site audits (@seekersight cut load times 40% via a Lighthouse loop), and multi-model orchestration before the deadline. The most repeated pattern of the day was outsourcing visual thinking to GPT-Image: @tounano showed Claude prompts GPT-Image better than GPT does, iterate visually on cheap mocks, then let any capable model do the 'monkey work' of turning screens into code or clean SVGs. @jcartu paired Fable's ADR-quality architectural reasoning with local GLM 5.2 inference at 90-110 tps on 4 RTX Pro 6000 Blackwells for the execution loop, sparking DGX Station vs custom rig debate. @tounano also surfaced a tmux workflow where one LLM observes another's live token stream to distill portable skills across Claude, Codex, and Gemini. Copywriting stayed contentious — @thewildzeno argued longform training on Halbert/Cialdini still produces cheesy TikTok scripts without human curation, while @mb29266 and @Guto_gouvea countered that few-shot examples per awareness stage produce VSLs that make readers cry. On security, @watchmedropship shared byte-level MCP metadata scanning and hash-pinning to catch Unicode-hidden payloads and description rug-pulls. Meta-threads on graph-indexing TG into a searchable forum, coopetition (alpha decays fast, sharing compounds), and a contribution-scoring algo rounded out the day.

Topics

Grok 4.5 shipped with 500k context (1M coming), tighter tool chains, and strong agent behavior — testers call it a generational leap. Anthropic extended Fable 5 access through July 12 (up to 50% of weekly usage), read as a hedge against Thursday's GPT-5.6 Sol launch. @seekersight cut site load times 40% via a Fable Lighthouse audit loop.

@jcartu detailed using Fable for high-level ADRs and scaffolding, then executing within-spec code on local hardware running an NF3 hybrid-quant GLM 5.2 at 90-110 tps across 4 RTX Pro 6000 Blackwells. Discussion covered DGX Station ($100k-$200k) vs custom rigs, plus @ltv_1's Fable-as-orchestrator setup with Codex/Opus/Sonnet workers.

@tounano evangelized outsourcing visual thinking to GPT-Image: Claude prompts image gen (better than GPT does), iterate on cheap mocks, then any model converts screens to code as taste-free 'monkey work'. Also produced a deterministic image→SVG pipeline for UI icons Codex can't match, and @MiamiMega open-sourced a SKILL.md version — ~100x more token-efficient than ingesting mockups.

@thewildzeno shared that training Opus on Halbert/Cialdini/Kennedy still produced cheesy TikTok scripts, arguing you need human copy expertise to grade output. @Guto_gouvea and @mb29266 countered with 'persuasion DNA' extraction and few-shot examples per awareness stage — longform VSLs and advertorials landing hard, TikTok short-form still the weak spot. @tounano's tmux trick (one LLM observes another's live tokens) distills portable skills across harnesses.

@watchmedropship shared four hardening moves: byte-level scanning + hash-pinning MCP tool metadata against Unicode payloads, a cron determinism scanner (59→2 µ$/item), cautious LLM-as-verifier testing, and click/scrape injection audits. @mikeconner posted a starter .claudeignore blocking node_modules/backups/media for instant token savings. @tounano proposed a graph DB where TG messages join multiple 'colonies' to preserve community knowledge.

Key Takeaways

  • Grok 4.5 is a legit generational leap — 500k context, faster tool chains, cleaner agent behavior; worth testing before committing to GPT-5.6 Sol on Thursday.
  • Pair Fable's ADR-quality architectural reasoning with cheap local inference (GLM 5.2 on RTX Pro 6000s at 90-110 tps) for the execution loop.
  • Claude prompts GPT-Image better than GPT does — iterate visually on cheap mocks, then let any model convert screens to code as taste-free work.
  • Harden MCP servers now: byte-scan tool metadata for Unicode payloads and hash-pin descriptions to detect post-approval mutation.
  • Alpha decays in weeks — sharing with trusted peers has higher EV than hoarding, and a .claudeignore at repo root instantly cuts context bloat.

Hot Threads

@geiltstarted

Grok 4.5 first impressions and multi-device orchestration setup

22 replies5 participants
@tounanostarted

Graph-DB indexing to turn TG chats into a searchable forum layer

25 replies6 participants
@jcartustarted

Fable ADRs + local GLM 5.2 inference rig economics

18 replies6 participants

Linked Items