Summary
The morning slot's public scrape and curated reposts came back empty, so everything here comes from the home-feed capture, which covers the US evening of 09-26. The strongest signal is economic rather than technical: one thread walks through the audited financials of MiniMax and Zhipu, the only listed pure inference companies, and another prices a "neolab" at $125M to $150M for 1,000 GB300s. The biggest cluster is again the decision-model wave, about fifteen posts on Jev and its clones. The substantive ones are jevgrep (a code-search CLI claiming 40% lower coding-agent cost on SWE-bench), RSI-Jev v2.0 (a 2B model improved by an agent team in about 15 hours), and Autorubric adding support for decision models; the rest are blueprint threads with no numbers. Agent research added two useful results: PRIMESCIENTIST, which spends a research budget across competing plans and uses half the attempts, and RSIAgent, which shows practicing broad before deep lifts success from 56.5% to 74.5%. Theo's "smart and dumb are separate axes" essay was the most discussed opinion post. Skip the "Anthropic, OpenAI and SpaceX mapped 40 billion data points" thread; it links nothing.
Posts
The economics of an inference company (@RMladek). Inference providers buy GPU time and sell tokens, while token prices fall and GPU prices rise. MiniMax and Zhipu (the Chinese labs behind the MiniMax and GLM models) listed in Hong Kong in January, so there are finally audited numbers. One went from losing money on every token to a 24.6% margin in 18 months; the other lost 75% of its OpenRouter volume in ten weeks. The wiki's 2026-09-23 digest already covered the margin number; this thread is the fullest walk-through.
Decision-model tools with numbers (cluster of 4) (@dzhng · jevgrep · @GaoShanghua · @deliprao · @karthikkalyan90). jevgrep is a CLI that lets a coding agent find files by asking what code does, using Jev (TypeSafe's small decision model) for context collection; it claims 40% lower coding-agent cost, verified on SWE-bench. RSI-Jev v2.0 is a 2B model an agent team improved with 50 experiments in about 15 hours: 0.79 versus Jev 1.13's 0.73 on typed decisions, and task gains did not transfer but calibration did. Autorubric 1.6.0 now grades with decision models as well as LLMs. A retrieval engineer argues decision models will make long-tail retrieval cheap by replacing parts of the vector-search stack. See the routing concept page.
Decision-model hype (cluster of 8, skip) (@marfinxx · @Mnilax · @hanakoxbt · @Av1dlive · @Dontgiveup_26 · @xmyttle · @mirku21 · @0xCodila). Ten-step blueprints, "97% fewer LLM calls" local runners, a Doom fight, and a "40 billion data points" thread. None links a benchmark. One useful fact: NanoJev, a 0.6B open replica, scored 128 of 128 on ViZDoom basic aiming against Jev's 56 (@rvaniaaaa).
Converting Qwen3 into a decision model, live (@neural_avb). A five-hour stream building a Jev-like model from Qwen3, including how to edit attention masks and position ids so the model's choice does not depend on the order options are listed. That is the same order-invariance trick Gero-4B uses.
PRIMESCIENTIST: spend the research budget like a portfolio (@rohanpaul_ai). Research agents usually push one idea until the budget runs out. PRIMESCIENTIST keeps competing executable plans in a tree, updates each branch's value from experiment results, explores while budget is plentiful and concentrates on the strongest branches as it shrinks. On 12 FIRE-Bench tasks: 10.3% higher reward than AutoResearch with 50.6% fewer attempts at the same token budget, and fewer attempts on 23 of 24 tasks overall.
RSIAgent: practice broad, then deep (@rohanpaul_ai). To adapt an agent to an unfamiliar app without retraining, a curriculum agent invents practice tasks, an actor solves them in code, and a separate verifier decides what is saved. Agents that practiced broadly before the hard cases averaged 74.54% on four tasks, versus 56.50% going straight to the hard cases.
Smart and dumb are two axes (@theo). Theo argues a model can be very capable and very error-prone at once. His examples: Gemini models know a great deal but do strange things at work, and Fable 5.1 is less capable than GPT-6 Astra but makes far fewer senseless moves. Elvis Saravia agreed it matches how he tests models.
Agent memory: optimize the right metric (@henrytdowling · @dair_ai). Stash's founder argues memory systems should hill-climb a better metric rather than stack machinery on the consensus one, citing Wispr Flow optimizing "user edit rate" instead of word error rate. dair.ai amplified a Salesforce paper finding agents should store raw trajectories rather than only summaries.
Minimal harnesses and multi-agent coding (cluster of 3) (@omarsar0 · @JoshARosen · @rvaniaaaa). An MIT CSAIL paper argues for keeping the agent harness minimal; an X article describes the end of the single-agent coding session as agents delegate to sub-agents in isolated environments; and a thread on a 535,000-line rewrite in eleven days for $165,000 credits the task graph (one job per agent, edges only where data flows), not the model. Related: the harness engineering concept page.
Learning resources (cluster of 3) (@attharrva15 · @0x0SojalSec · @_avichawla). An X article on vLLM's architecture (body not captured, click through), a free 16-hour series building a DeepSeek-style model (multi-head latent attention, grouped-query attention, MoE), and a repost of the interview question on why a top-2 MoE (mixture-of-experts) can be slower than a dense layer despite fewer FLOPs: cross-GPU expert traffic.
Deep networks as kernel machines (@pmddomingos · paper). Pedro Domingos re-posts his 2020 result that networks trained by gradient descent are approximately kernel machines, which memorize training data and predict by similarity, as a reply to a new finding he calls unsurprising.
Dream-RSI (@Sa4d_k1 · paper). An Arabic-language explainer of a Google DeepMind, Maryland and Virginia paper arguing recursive self-improvement is bottlenecked by exploration, and that the exploration strategy itself should evolve with the search space.
Contrastive World Models (@shumpeiMaxwell · paper). Replacing Dreamer's pixel decoder with a contrastive objective learns the task-relevant dynamics and ignores visual distractors like moving backgrounds.
Skip: a Palantir agent-stack explainer, a free ElevenLabs clone promo, Karpathy and Andrew Ng lecture-summary threads, a NeurIPS creativity acceptance, and an Arabic "Google automates the PhD" post.