Issue · 2026-09-21

Most AI news is noise.Here is the rest.

DIGEST

Jev left the waitlist, and the ecosystem with it

Jev left the waitlist today (typesafeai) and now sits on OpenRouter beta and Vercel AI Gateway as typesafe-ai/jev. TypeSafe claims 193x faster and 444x cheaper than Fable 5.1 or Astra on classification (DataChaz).
Compaction idea to steal: score every tool call and drop the irrelevant ones instead of summarising (tamarajtran). An independent eval found it degenerated into a rule a free script could run (Teknium).
Small pieces landing: a skill pack for typed decide/tool/evaluate loops in Codex or Cursor (LLMpsycho), a per-turn Jev verifier that makes goal-checking cheap (omarsar0), and the same trick for RL grading (Vtrivedy10).
Or own the primitive: Kev-0.5B runs on a MacBook Pro (jaredpalmer), now grown to 0.6B/4B/8B on Qwen3 with a public recipe (jaredpalmer); an MLP on a 4B Qwen does the same job (justALEXWORTEGA).

DIGEST

Shipped: DeepSeek-V4.1-Flash, 552B MoE, 1M context, MIT weights, beating the larger V4-Pro on most agentic and coding benchmarks (HuggingPapers, paddix).
Cheap because the KV cache compresses to ~890 bytes per token, a quarter of V4-Flash, with 8B parameters active on prefill and 16B on decode (getmeosu, calebfoundry); the paper's chart explains the 91% cost gap (KanikaBK).
V5 is rumour only — beats Astra, matches Fable 5.1, open weights, cheaper (Mr_Salio) — with no model card and no benchmark sheet (vepsi__).

DIGEST

Hit rate comes from KV prefix caching plus append-only sessions (stretchcloud); one provider reports 94% on DeepSeek traffic (vipulved); a claimed "99%" means the claimant is reselling DeepSeek, the only one that hits it (thdxr).
Same builder on the claims around it: reproducing DeepSeek-grade inference took months, ~$100M and the right ten people (thdxr), and "open source is overtaking frontier" charts come from routers that see too little traffic to mean anything (thdxr).
The trap is walking away: one uncached request at 500k context costs over $5 — the cache expires after 1h on Claude, 30m on Codex (kunchenguid). Two fixes: strip a coding agent to read/inspect/write, under 2k startup tokens, 200k window (alexgetmancom); in Pi, switch on cache-miss notices in /settings (Agents365_ai).

ORCHESTRATION

SoL-Pi runs auto-research loops at the harness layer, cuts token traffic nearly in half and still matches a hand-tuned baseline (omarsar0) — and ships as a Pi coding-agent extension, not a standalone system (agenticgirl), with 4 of 152 proposed mechanisms surviving (GitGem).
Framing of the week: Google's nine-page playbook reduces to Agent = Model + Harness, same model and benchmark, different result (AnnatarXBT); harness-task fit dominates model-harness fit (Vtrivedy10); it is not bigger CLAUDE.md files but designing the environment (simas_ch); and a short from-scratch guide you can feed an agent (omarsar0).
The layer is becoming company infrastructure: YC open-sourced its internal multi-agent harness QM (stretchcloud), one team moved onto Nous Research's Hermes and costed out what it took (JacquelineSYC19), and one builder is hillclimbing a Jev harness for code search in the open (dexhorthy).
Andrew Ng says prompting dies within six months, replaced by loops-and-graphs harnesses (0xwhrrari). Against that: agent-assisted planning was harder than pen and paper, with distracting questions and early conclusions (mattpocockuk).

DIGEST

0.86.0 adds mid-conversation system messages, dynamic tools without losing the KV cache, Anthropic cache warming and faster resume (pidotdev); pi-observational-memory is the most common crash cause on it, so keep that extension disabled until the upstream fix lands (pidotdev).
pi-jev v0.4.0 gates diffs and workflow judgments through Jev inside Pi (NotThatTheo); 0.86.1 retains the system prompt better, and a 9B local model at ~5GB with vision is workable (str8bored).

DIGEST

Rumoured for next week: GPT-6 Sol on Tuesday, cheaper and stronger than 6 Astra (imjustnewatai); Opus 5.5 in stealth testing as claude-wafer-eap, aimed at Astra and Fable 5.1 level (pankajkumar_dev); Sonnet 5.2, Fable 5.2 and Gemini 4 Pro already spotted in testing (kimmonismus). For today's choice: Fable versus Astra for app development (twostraws).