Marktechpost Co-Editor Jean-Marc Mommessin recently sat down for an exclusive talk with Eddie Siegel, CTO/Co-Founder at Ode with Anthropic, about their work in AI. The discussion dives deep into Ode's core AI technology and what lies ahead for AI-driven solutions. piped.video/watch?v=4o7PCGd6… @OdeDevs

We tried the SpeakON's MagSafe AI Voice Button and its really cool! It has its Own Microphone: Turning Your Voice into Polished Communication, and Action across Apps Voice input on phones has been solved for years. What has not been solved is the output. Speak into most dictation tools and you get back exactly what you said, fillers and false starts included, in a note you then have to clean up and move somewhere else. SpeakON attacks that gap with hardware: a 25 g magnetic button that snaps to the back of an iPhone, carries its own microphone, and writes finished text straight into whatever app is already open. Read our full analysis: marktechpost.com/2026/09/22/… Try it here: speakon.sjv.io/Gbd2EL @SpeakON_Global

Nokia AI Research Open-Sources AnyJev: A Training-Free Layer That Turns Any Open LLM Into a Calibrated Decision Model It reads typed answers straight from the model's next-token distribution. It removes position and label bias without any training. ▪️ Apache-2.0, pip install "anyjev[hf]" ▪️ Choice, yes/no and score questions, nothing generated ▪️ Hugging Face transformers and vLLM backends ▪️ L0 needs zero labels, L1 needs 100 to 500 ▪️ Order-flip rate 0.230 to 0.073 at L0 ▪️ Calibration error 0.240 to 0.095 at L1 ▪️ Auto-decidable at 5% error: 7.7% to 52.0% at L1 ▪️ Tested on Qwen3-8B, BANKING77 20-way, 300 test items Accuracy moves 6 points, but the traffic you can safely automate grows 6.8x. Full analysis: marktechpost.com/2026/09/23/… Repo: github.com/nokia-applied-res… @nokia

NVIDIA Releases Nemotron 3 Diarization: A 100M-Parameter Open-Weight Model That Tracks 8 Speakers in Real Time It answers "who spoke when" in a conversation, even when people talk over each other. One open-weight checkpoint handles both offline audio and real-time streaming. ✅ 100M parameters, open weights on Hugging Face ✅ OpenMDW-1.1 license, commercial use allowed ✅ Tracks up to 8 speakers, overlap included ✅ 14.72% DER on Diarization-Bench vs 19.3% runner-up ✅ 4 streaming presets, from 30.4 s down to 0.32 s ✅ 41.0% average relative DER cut vs Streaming Sortformer ✅ 15,113x batched RTFx on RTX PRO 5000, 30.4 s preset ✅ Runs through NVIDIA NeMo on Linux GPUs Full breakdown + interactive explainer: marktechpost.com/2026/09/23/… Model: huggingface.co/nvidia/Nemotr… NVIDIA blog: huggingface.co/blog/nvidia/n… Live demo: huggingface.co/spaces/nvidia… @NVIDIAAI

We have created a Coding Guide with open notebook to try TypeSafe AI's Jev: Typed Decisions, Calibrated Confidence, and Speculative Fan-Out with a System One Model TypeSafe launched Jev last week. It's their first "System One" model, and it doesn't write text at all. You send it program state (a string, array, or JSON object) plus a set of typed questions. It returns 1 of 3 things per question: → Choice: picks a label from options you define, with a probability for every label → Score: places the state on an ordered rubric, and the result can land between levels → Noul: a single probability that a statement is true No parsing and no "please respond in valid JSON." I wrote a full coding walkthrough with the official Python SDK. The patterns that I think matter most: 1. Confidence-gated routing. Classify intent, then act only if confidence clears a bar that rises with the stakes: 0.5 to read a balance, 0.9 to close an account. The thresholds are plain Python values, so you can review, version and test them. They're not buried in a prompt. 2. Speculative fan-out. Questions in one request can't see each other, so you can ask everything up front, including questions that only matter on one branch. We put 10 questions in 1 call and asked each again in its own call. The state gets sent once instead of 10 times, and each answer should match whether it's asked alone or in a batch. The notebook checks that. 3. Function calling as closed sets. One Choice picks the tool (including an explicit "none" option). One Choice per argument is asked in the same request. Your code only executes with validated, enumerated values. 4. Keep the weights in code. For ranking, ask the model for per-dimension Scores and store them. Ranking is then just arithmetic. If you change the weights, the list re-ranks with zero extra inference calls. The honest limits: → Jev can't count reliably within a single question. The workaround is 1 Noul per item, with the sum done in code. → It's the wrong tool for chat, codegen or anything that needs an explanation. → It's a closed API in early access. The speed and cost numbers circulating are TypeSafe's own and haven't been independently benchmarked. → You still have to evaluate your questions, criteria and thresholds on your own data before trusting them with real actions. The notebook also covers Pydantic response models, async fan-out with asyncio, retry policies, typed errors, and a token ledger that prices the whole run. Curious whether anyone here has tried this against a small fine-tuned classifier or logprob-based routing on a regular LLM. That feels like the real comparison. Full tutorial with code: marktechpost.com/2026/09/23/… Docs: docs.typesafe.ai Notebook: github.com/MARKTECHPOST-AI-M…

Marktechpost Co-Editor Jean-Marc Mommessin recently sat down for an exclusive talk with Eddie Siegel, CTO/Co-Founder at Ode with Anthropic, about their work in AI. The discussion dives deep into Ode's core AI technology and what lies ahead for AI-driven solutions. piped.video/watch?v=4o7PCGd6… @OdeDevs

The web now has a subscribe button. Introducing TinyFish Monitor. Free for the next 72 hours! Point it at a page or a topic. Write what you care about in plain English. Monitor does the checking and judging. You or your agent only hear about it when it happens.

NVIDIA used AI-run research loops to cut a coding agent's token traffic by up to 49%. They Introduced SoL-Pi: An MIT-licensed extension for the Pi coding agent. It packages 4 mechanisms that an AI discovered by testing harness changes across 535 environments. • MIT license, opt-in, installs on unmodified Pi 0.85.1 • 152 proposed directions, 4 mechanisms retained • 44.7% to 49.0% fewer tokens vs Pi on EdgeBench • About 33% lower API cost vs Pi • 93.7% to 94.3% of Pi's EdgeBench score retained • 50.0% to 54.3% cheaper than native harnesses on EdgeBench • Built on GPT-5.6 Sol, transferred to Opus 5 unchanged • Mechanisms: Action Fusion, ObservationPack, Online Context Compact, Evidence-Preserving Reducer Full analysis: marktechpost.com/2026/09/21/… Paper: arxiv.org/abs/2609.20519 GitHub: github.com/NVlabs/SoL-Pi Blog: nvlabs.github.io/SoL-Pi/ @NVIDIAAI @NVIDIARobotics @nvidia

Grok 4.7 just dropped at $2/$6 per million tokens. Fable 5.1 Max costs $10/$50. Grok 4.7 still beats it on EEBench (64.0% vs 56.4%) and Harvey Legal (19.6% vs 6.7%). Same price as Grok 4.6. Bigger base model. Full analysis: marktechpost.com/2026/09/21/…