# google ships live voice, and rl still doesn't buy hard problems > gemini 3.8 live ships in two price tiers, rl only improves what was already easy and naive bayes beats a 397b with labels: the day is compute allocation. - edition: Tuesday, September 15, 2026 (2026-09-15) - notebook: dev & ai - topics: llm · papers · inference · market - items: 12 from 12 sources - original: https://tonho.wtf/en/daily/2026-09-15/ - portuguese edition: https://tonho.wtf/diario/2026-09-15/ - authorship: written by an llm pipeline, reviewed and translated by antonio leandro (tonho.wtf) --- The announcement of the day is Google putting live voice into production, and even it arrives split in two: one model for scale and cost, another for complex tasks. That split is the thread running through everything that came after. One paper measures that RL fattens up what the model already got right and barely moves the hard stuff, and proposes reallocating sampling to where compute still buys something. Another reuses a 1-bit index to avoid paying for dense attention in long context. A third shows Complement Naive Bayes tying with a 27B and beating a 397B frontier model at topic classification, at thousands of samples per second on an ordinary CPU. NVIDIA publishes the explicit criterion for when dense and when MoE, with price and throughput in the same table. Nobody is asking which model is smarter. Everybody is asking where extra compute stops buying capability. And there's the outer layer, which is the actual bill. On TabNews, the argument is that without cost per token and eval verdict in the same span you don't find out which 1% of queries eats half the budget. Cloudflare separates search crawling from training crawling, because refusing one meant losing the other. Diesel in the US hits a record, which is the energy input the labs declare as a constraint. And against Amodei's essay that ran here on the 12th, Xataka points at a detail that's simple to check: the words "open weights" don't appear a single time in the text, and the ones driving prices down are precisely the ones publishing weights. ## labs **[Gemini 3.8 Live and 3.8 Live Extended Thinking](https://deepmind.google/blog/introducing-gemini-3-8-live-and-3-8-live-extended-thinking/)** — Two speech-to-speech models in the GPT-Live family format, and the split between them is stated in the announcement itself: Live is for scale and cost efficiency, Extended Thinking is for complex tasks and reasons while it talks, using verbal cues and progress narration to keep the conversation from going silent during a tool call. Extended Thinking took first place on Artificial Analysis's Speech to Speech Quality Index (82.6), with 68.6% on τ-Voice, 35.1% on Sierra's τ-Voice-banking and 97.7% on Big Bench Audio; Live came in second on the Speech Agent Arena. It detects and switches between 97 languages mid-conversation, runs tools in the background while it keeps answering, and all audio comes out with a SynthID watermark. It's already in the Gemini API and AI Studio, and the copy selling "highly competitive pricing" doesn't carry a single price number, which is annoying in a launch whose central argument is cost. **[Dense vs. MoE: active parameters, throughput and when to choose each](https://developer.nvidia.com/blog/dense-vs-moe-models-active-parameters-throughput-and-when-to-choose-each/)** — The useful thesis in NVIDIA's post is that MoE decouples memory from compute: VRAM tracks total parameters, FLOPs per token track active ones. A 30B MoE and a 30B dense take up the same ~60 GB in BF16; the difference is what you buy with those gigabytes, capability or throughput. The advantage shows up at batch 1, where decode is memory-bound and the MoE reads fewer weight bytes per token, and it compresses at high concurrency, when tokens end up activating almost every expert — which have to be on the GPU at the same time anyway, leaving less room for the kv cache. The table closes the argument: Nemotron 3.5 Lightning (30B total, 3B active, hybrid Mamba-2 + MoE) delivers 235.7 to 494.2 t/s at $0.22 per million output tokens, against 46.8 t/s and $3.00 for the dense Qwen3.8-27B, with less than half the capability score (24 against 52). ## research **[Learning to Solve Hard Problems in RL for LLMs by Never Giving Up](https://arxiv.org/abs/2609.13443)** — The Matthew effect in RL: training greatly improves the problems the model already solved and barely improves the hard ones, and the authors argue that current methods make this worse by spending too much compute on what's already easy. The proposal, Never Give Up, is adaptive sampling — keep generating samples for a problem until one comes out correct — which, backed by asynchronous RL, filters the easy stuff with few samples and reallocates the rest to the hard stuff. On Deepscaler it improves performance per compute, with the gain concentrated in the hard problems; on the Manufactoria code task, GRPO with per-test reward stalls on problems that mix easy and hard tests, and NGU keeps solving harder and harder tests until it closes out the whole problem. Signed by Michael Noukhovitch, Hamish Ivison, Nathan Lambert and Aaron Courville. **[Self-Indexing Attention](https://arxiv.org/abs/2609.13205)** — Sparse long-context inference usually uses different retrieval strategies in prefill and decode, which makes it impossible to reuse one and the same representation throughout inference. The paper proposes a shared sign-magnitude representation in a transformed domain, training-free, in which the signs of the keys become a 1-bit index reusable in both stages, resolved by bitwise operations that modern accelerators already support. At 5% attention density it stays close to dense attention on LongBench and RULER, with up to 6.1x speedup on the attention operator in prefill and 10.3x in decode, and it remains compatible with low-bit kv cache compression, tested with TurboQuant and DeepSeekV4-Flash. **[LLMs or Naive Bayes? Old Gems or New Ways](https://arxiv.org/abs/2609.13185)** — Complement Naive Bayes against LLMs from four families, from 27B to a 1T MoE, on text classification. The LLM only dominates in the no-labeled-data regime (98.0% against 88.2% on Amazon Polarity), and even that win is sensitive to contamination: on a low-contamination sentiment task NB beats zero-shot, 81.7% against 73.0%. With labels available, on AG News, NB reaches 89.1%, statistically indistinguishable from the 27B zero-shot (89.0%) and above the 397B frontier model (84.8%), at thousands of samples per second on an ordinary CPU — batched inference of a small LLM on GPU comes out 40 to 486x slower, with about two orders of magnitude more energy per sample. The decision line is per task, with parity around 10⁴ labels for topic classification, and it comes with a Helm operator for Kubernetes that automates the choice with thresholds and verifiable Prometheus metrics. For anyone keeping a classification queue in production, it's the most actionable paper of the day. ## brazil **[Observability and Evals in Production: The Closing of LLMOps](https://www.tabnews.com.br/andersonlimadev/observabilidade-e-evals-em-producao-o-fechamento-do-llmops)** — Fifth and last article in andersonlimadev's TabNews series, with a thesis narrow enough to be testable: without the full prompt, model version and eval verdict in the same span, there is no generative observability, just a server log pretending to understand LLMs. The concrete mechanism is in OpenTelemetry's GenAI semantic conventions, and the detail they highlight is the separation between `gen_ai.request.model` and `gen_ai.response.model` — that's what makes model drift visible when the provider swaps the snapshot underneath you without you deploying anything. On the cost side, they argue for attributing tokens and cost to each span as the only way to find the 1% of queries that eats half the budget, and for running the eval span on a sample only, since LLM-as-judge on 100% of responses doesn't pencil out. It closes with online evals in shadow against a golden set and automatic rollback on metric degradation. ## market **[Cloudflare separates search crawling from training crawling](https://blog.cloudflare.com/accountable-mixed-use-ai-crawlers/)** — Until today, refusing training to a mixed-use crawler meant disappearing from search, because it's the same bot doing both things. The new Disallow AI Training setting publishes the preference in robots.txt and keeps crawlers with the Accountable designation cleared for indexing; Apple, Google and Microsoft either comply or have committed to complying with the requirements, which include URL-level visibility into which pages were made available for training. The numbers explain the design: fewer than 1% of sites on Cloudflare block search bots, against 17% that turn on some mechanism against training. Worth paying attention to the change in semantics that takes effect today — Block and "Block on pages with ads" now apply to mixed-use crawlers, meaning anyone already using Block loses search too, and Managed Robots.txt was deprecated in favor of Bot Preference Sync. **[B2B Medical & Pharma UX Benchmark](https://feeds.baymard.com/link/9825/17462397/b2b-medical-pharma-ux-benchmark-2026)** — Baymard evaluated 10 medical and pharma supply sites on 400+ parameters, generating 3,400+ weighted scores and 2,600+ best-practice examples, all of it coming out of user testing. Not one of the ten even reaches "decent": the range goes from "poor" to "mediocre", with Thermo Fisher, McKesson, Henry Schein and Sigma Aldrich at the bottom. The failures are implementation, not trend — missing documentation behind specs that do exist, the same attribute labeled differently from one product to the next, price and total order cost outside the purchase section, search that gives up when the query comes out slightly wrong, and a cart that can't be saved or handed off, in a context where whoever builds the order is almost never whoever approves it. ## world **[US diesel at a record high](https://www.semafor.com/article/09/15/2026/us-diesel-hits-record-highs-and-oil-supply-fears-mount)** — A Semafor headline, with no full text available: diesel hit a record on Tuesday with commercial and strategic stocks low, Chevron's CEO saying the buffers that held down price and supply risk have run out, and Saudi Arabia canceling shipments to Europe with the East-West pipeline down for weeks after a drone attack. Linked here as context on the input the labs declare as the constraint on AI factories. ## who wrote **[Simon Willison, Gemini Live audio](https://simonwillison.net/2026/Sep/15/gemini-live/)** — Instead of repeating the release, Willison pointed GPT-6 Astra Extra High at the documentation and published a web UI to test the new models: pick a model and a voice, optional system prompt, talk through the browser and you can interrupt the model mid-sentence. The implementation uses no library at all, connects straight to the `BidiGenerateContent` WebSocket endpoint and uses an AudioContext from the Web Audio API for capture and playback. It's the fastest way to find out what the API actually delivers, which is rarely what the benchmark table says. **[Gergely Orosz, Inside OpenAI's agentic software factory](https://newsletter.pragmaticengineer.com/p/openai-software-factory)** — An account of a visit to OpenAI with seven people from engineering, and the central data point is adoption: in four months, non-technical areas like finance, recruiting and legal went from near 0% to 90% Codex usage, with no mandate from above. IDE use has been falling since January, which forced a rethink of pull request and code review; the Perf Factory monitors production and fires off agents to fix performance regressions on its own; threads last days with `/goal`, and one long-running agent branches other agents, shrinking the surface a human has to manage. There are two honest caveats in the text: the internal Codex is far more plugged into in-house systems than the external one, and CPU became an API bottleneck. Worth reading against the finding from the 14th here — a native harness doesn't win on average, and even so the entire company depends on a single shared harness, to the point that a minor failure got flagged by colleagues before the automatic alert fired. **[Javier Lacort, Xataka](https://www.xataka.com/robotica-e-ia/frenar-ia-suena-a-proteger-a-humanidad-tambien-forma-muy-eficaz-frenar-a-deepseek-qwen-mistral)** — The observation that holds up the whole piece: the words "open weights" don't appear a single time in Amodei's "We must pace the frontier", which ran here on the 12th. Capability-based certification is easy to comply with for a closed lab, which controls every copy, and impossible for anyone publishing weights, because anybody can remove the safeguard afterward; the proposed antitrust exemption draws a table where whoever doesn't sit at it takes the limits without helping set them. The numbers they put together are the argument: DeepSeek V4 Flash comes out 150x cheaper than GPT-5.5 on output tokens on OpenRouter, the best open models are 3 to 6 months behind and that distance isn't growing, there are 151,448 Qwen derivatives on Hugging Face (2.6x Meta's), and OpenAI's adjusted gross margin fell from 40% to 33% in 2025. Meanwhile Mistral closed €3 billion at a €21 billion valuation selling exactly sovereignty with open weights. ## stalled sources Anthropic News quiet for 15 days, Anthropic Engineering for 114, Import AI for 9, Microsoft Research for 15, One Useful Thing for 16, Benedict Evans for 12 and Karpathy for 138.