🚨 BREAKING: Super excited to announce we're deploying the world's fastest inference with @Cerebras. Talk to any developer and they're excited to build with 20x faster AI... the problem is there's almost no compute available. We're here to solve that. Using GPUs for prefill - it's now more affordable than ever too. Thank you to the whole Cerebras team and excited to grow this partnership. First tokens live Q127 🚀

🚨 BREAKING: Super excited to announce we're deploying the world's fastest inference with @Cerebras. Talk to any developer and they're excited to build with 20x faster AI... the problem is there's almost no compute available. We're here to solve that. Using GPUs for prefill - it's now more affordable than ever too. Thank you to the whole Cerebras team and excited to grow this partnership. First tokens live Q127 🚀

🚨 BREAKING: Super excited to announce we're deploying the world's fastest inference with @Cerebras. Talk to any developer and they're excited to build with 20x faster AI... the problem is there's almost no compute available. We're here to solve that. Using Nvidia GPUs for prefill - it's now more affordable than ever too. Thank you to the whole Cerebras team and excited to grow this partnership. First tokens live Q127 🚀

YC's @garrytan says hard tech startups make up a quarter of the newest 200-company Demo Day batch, up 40x since their low point ~4 years ago: "It's 200 companies presenting today. About a quarter of them are hard tech. That's up 40x since our low bar, maybe 3-5 years ago."

One cool technique to improve latency in LLMs is to have a small, fast model generate tokens and a larger, slower model review the output in one pass. It'd be like having a junior employee write a report, and then having a senior employee review it in one go before it gets sent to a client. This is called "speculative decoding" and lets you generate faster while maintaining the output quality. Great write-up from the @DecagonAI team on how we've been doing this:

What changes when you're building software for agents not for humans? “The experience is going to be defined by the slowest point in that chain. The number one requirement for agentic query patterns is low latency, because they are executing dozens of SQL queries simultaneously across all these different systems. The most important requirement is the unpredictability of those query patterns, the responsiveness, and the fact that they are much more exploratory than a traditional human report or query.” @ceo_clickhouse Love to hear your thoughts @glcst @kiwicopple @bernhardsson @ashashutosh

Announcing the Artificial Analysis Pronunciation Robustness benchmark, measuring how reliably Text to Speech models say challenging text correctly - Google Gemini 3.1 Flash TTS leads at 88.1%, followed closely by SpaceXAI TTS at 87.6% and ElevenLabs Eleven v3 at 85.6% Existing Text to Speech (TTS) evaluations, including our TTS Arena, capture overall listener preference (e.g., how natural a voice sounds), and Word Error Rate (WER) checks whether the right words come out. Pronunciation Robustness adds a view of whether those words are said correctly (e.g., reading “St.” as “Saint” and “Street” in “St. Mary’s is on Church St.”) - this is critical for production voice agents, which need to get account details, names, currency amounts and more right to be trusted by users, at the low latency that conversational experiences demand. Overview of Pronunciation Robustness Each model generates audio for 454 sentences containing 701 target words or phrases, across four categories: 1. Contextually appropriate (words read differently depending on context, e.g., a wound that is bandaged vs. a bandage that is wound) 2. Expanding shorthand (numbers, dates, units and notation read out naturally, e.g., 6'2", Chapter XVII or 1 tsp of sugar) 3. Preserving exact sequences (codes, paths, emails and identifiers spoken exactly, e.g., .env.local or a.chen@ucsf.edu) 4. Standalone terms (brand, place and technical names, e.g., Arkansas, façade or genre) Sentences are sent as written, with no normalization on our side beyond each model's default setting. Screened human listeners judge whether each target was pronounced correctly against accepted pronunciations set in advance, excluding anyone who fails an attention check. The score is the share of correct judgements, excluding "I could not tell" answers, and we publish a model once 95% of targets have three or more approved listeners. Key results: ➤ Overall leaders: @GoogleDeepMind’s Gemini 3.1 Flash TTS leads at 88.1%, followed by @SpaceXAI’s TTS at 87.6%, @ElevenLabs’s Eleven v3 at 85.6%, v3 Conversational at 84.8% and @Alibaba’s Qwen-Audio-3.0-TTS-Plus at 81.6%. Gemini 3.1 Flash TTS leads Contextually appropriate and Expanding shorthand, SpaceXAI TTS leads Preserving exact sequences, and Qwen-Audio-3.0-TTS-Plus leads Standalone terms. ➤ Hardest categories: Expanding shorthand (62.4%) and Preserving exact sequences (62.9%) trail Contextually appropriate (86.2%) and Standalone terms (86.1%). We expect Expanding shorthand and Preserving exact sequences scores to rise considerably with normalized input text, especially for models without normalization on by default. ➤ Preference ≠ pronunciation robustness: The most preferred voices aren't always the most accurate - Sonic 3.6 ranks #1 on the Provider Voice Arena at 1276 Elo but #11 on pronunciation robustness at 74.5%, while Gemini 3.1 Flash TTS ranks #9 on the Arena at 1201 Elo and #1 on pronunciation robustness at 88.1%. See more details below ⬇️