How Souffleur turns AI voice conversations into glanceable topic cards
Souffleur is a research prototype that displays voice chats as topics, decisions and short gists. Its design shows what a compact overview can leave out.
AI product releases, demos, and workflows. Find out what you can try, what access or setup it requires, and what evidence you have for choosing a tool.
Browse all articlesLook at the inputs, the steps people still handle, and what comes out. These stories examine tools for conversation, writing, design, and enterprise work.
Souffleur is a research prototype that displays voice chats as topics, decisions and short gists. Its design shows what a compact overview can leave out.
Maike's proposed writing-tutor architecture links essay diagnosis, targeted questions and revision planning to edits students write themselves.
OpenAI, Microsoft and TypeSafe's Jev all shipped models that answer with a probability over the options you supply instead of generating text, within a few weeks of each other. Here is what the class does, what is actually new, and how far the vendor numbers can be trusted.
Blueprints AI, the Santa Monica startup TechCrunch profiled on 12 October 2026, sells software that turns a sketch or a written description into construction documents and advertises them as permit-ready. No independent evidence shows an AI-generated drawing set passing plan check in any jurisdiction; the strongest municipal data covers AI pre-screening of applications, and liability stays with the licensed professional who seals the set.
Google says its enterprise agent runs under its own Workspace identity, logs every action under its own name, and can be paused by a per-project spend cap. Release status, retention and pricing are unstated, and every customer figure is vendor-supplied.
Uber is giving ChatGPT Work, OpenAI's agent running inside the company's ChatGPT Enterprise account, to more than 30,000 employees, after a 1,000-person pilot in which an Uber engineering vice president told TechCrunch that about 80% of adopters were non-engineers. Uber's own announcement confirms the scale and names no spend, no seat count and no measured result.
Check whether the weights are available, what the licence permits, and what the release says about hardware and runtime.
Mistral's improved Large 4 weights can be downloaded and run privately. Commercial rights depend on MaaS conditions, while serving needs vary by precision and workload.
Nvidia shipped Nemotron 3.5 Super VL and Bolt 74B-A9B on 15 October under a permissive Linux Foundation licence with no field-of-use limits, but part of the training mix stays private and local deployment is a single-stream affair on 24GB.
Recursive's Flint is a 40B-parameter mixture-of-experts coding model with downloadable Apache 2.0 weights, a documented architecture and dozens of community quantizations. The benchmark scores, throughput figures and the claim that Recursive's Eureka Machine built much of the model are the company's own; outside it, two benchmark figures have had a small-sample check, and none of the headline results has been reproduced.
Block says Dessa was trained from scratch on human conversation and runs on a MacBook. No member of the public can check either claim, and the page says nothing about whose voices were used.
Moonshot's Kimi K3.1 Pro scores 55 on the Artificial Analysis Intelligence Index, fourth overall and three points behind the leader. The model runs today in Kimi's products and API, but the full open weights are promised by 4 Nov 2026, and Moonshot says it removed cyber-related data from two training stages while intending to raise that capability over time.
When you compare tools, check model defaults, which plans can use them, and how the provider charges for a request.
GPT-6.1 Terra costs $1.00 per million input tokens and $5.00 per million output, but any prompt with more than 272K input tokens is billed at double the input and cache rates for the full request. A 280K-token prompt costs about twice a 270K one. Cache writes, Batch, Flex and Fast mode move the total again, regional processing adds an opt-in 10%, and fine-tuning is not available.
Meta released Muse Wonder and Muse Wonder Full on 12 October. Muse Wonder costs $4 per million input and $20 per million output tokens on the API; Muse Wonder Full comes only with a $50 or $200 a month Muse Code subscription. Meta's parity and MemoryBench claims rest on its own testing, and the one independent ranking found puts Muse Wonder Full fourth, behind Claude Opus 5.5.
Replit's CEO told TechCrunch Disrupt that the company is no longer working with Anthropic, and later clarified that OpenAI's models are now the default for Replit Agent. Insiders dispute the claim that the partnership ended, Anthropic has not commented, and Replit's own documentation never mentions a default model.
Google says OpenAI's GPT-6 Astra, GPT-6.1 Sol and GPT-6 Luna are in its model garden in private preview for select enterprise customers, transacted directly through Google Cloud Marketplace so usage lands on a single invoice and draws down existing cloud commitments. Zero Data Retention, Confidential Compute, availability in Gemini for business and in Google Antigravity are still promises, and the 'only hyperscaler with all three' line is Google's framing: 'leading' is undefined and the Claude leg is already generally available.
Arvind Krishna told IBM's 21 October earnings call that open-weight models are now about 90 percent as effective as frontier models at 10 to 25 percent of the token cost. That is an attributed spoken claim, no independent benchmark, and IBM's written results release carries none of the figures. We price it per workload against public list prices and name the evaluation that would settle it.
Pipecat’s BSD 2-Clause licence and provider choices are documented today. Deepgram’s Daily acquisition announcement adds promises about a Foundation and an integrated platform.
Google has announced staged Carbon access for cyber defenders. Independent index results, patching benchmarks and Fairwind policy offer different kinds of evidence.
