News, product updates, and insights from the Telli.sh team.
What are AGI and recursive self-improvement? Follow Google's RSI rumor back to the original X posts, examine the Amodei–Altman–Musk calls for pacing and Trump's China argument, and explore a faster, more hopeful future that still leaves people in control.
Memory, project rules, and the prompt you type today all influence an AI assistant. As models follow instructions more closely, old workarounds and vague permissions can become persistent problems. A practical guide to auditing what your assistant remembers, defining useful autonomy, and writing instructions you can actually test—with two templates to adapt.
A model's training run no longer tells the whole story of its capabilities. Research from Reflexion and Voyager to AlphaEvolve, test-time training and 2026 continual-learning experiments shows how AI systems can retain useful experience. What actually changes, where the evidence is strong, and why reliable improvement matters more than an endless loop.
GPT-Live-1 reaches the API at $0.05 a voice minute, three months after Google's 70-language live speech translation. What the published prices and benchmarks measure, what they don't, and why meeting records, shared ideas, and saved sources matter more as conversation gets easier.
Between September 1 and 3, 2026, Anthropic, Meta, Google and OpenAI all shipped frontier models, and CNBC gave the resulting exhaustion a name: model fatigue. The number of releases is the boring part. The interesting part is that every release creates an evaluation somebody downstream has to run, and the companies shipping them have no reason to help you skip it. A look at what a model swap actually costs, why cheaper per token is not cheaper per task, and why fully-managed software built for one narrow job is the only arrangement where somebody else can make the decision for you.
On September 3, 2026, NVIDIA agreed to acquire Hugging Face for $12,930,300,000 — its second-biggest deal ever, after $20 billion for Groq's assets. Hugging Face began in 2016 as a chatbot app for teenagers and ended up hosting 3 million models for 18 million developers. A close read of how it got there, why NVIDIA's two previous platform bets ended in opposite ways, and where the neutrality promise will actually be decided: not in a blog post, but in kernels, quantization formats and default deploy targets.
Clip turns a web page into a Telli.sh note with an AI-written title and a one-line summary, not just a saved link. It shipped behind an admin flag, so two analytics counters never fired once. That gate is gone: free accounts get 30 saves and 10 AI summaries a day, and version 1.18 of the browser extension adds a right-click Save to Telli.sh that authenticates with a token which can do nothing but save clips.
OpenAI shipped GPT-6 Astra on September 3, 2026 and called it the world's most intelligent model. Its own announcement tables tell a sharper story: enormous jumps on narrow tasks — Terminal-Bench Science 22.4% to 64.6%, SRE-Bench 55.9% to 88.0%, OSWorld at 47% less time per task — sitting on top of a neutral aggregate that moved three tenths of a point. Where the gains are real, where the asterisks are, and what a 1M-token text model still cannot do.
The Universal Commerce Protocol shipped on January 11, 2026 with Google, Shopify, Etsy, Wayfair, Target, Walmart and 20+ endorsing partners. Read the spec and the shopping is the boring part: /.well-known/ucp discovery, RFC 9421 message signatures, mandatory webhook signing, and an AP2 mandates extension whose entire job is making an agent's purchase non-repudiable. A close read of what UCP actually standardises — and the one thing it standardises that has nothing to do with buying.
Telli.sh: Every Translator 1.16 rebuilds the toolbar popup as a status view: a miniature of the page you are on, five labelled chips, and dots that go green only when the extension can confirm it is running. Every setting moved to a dedicated settings page with live skeleton previews, instant translation gained its own tab with a self-pruning 30-day history, and a notification bell now carries the Telli.sh blog in your browser's language across 15 locales.
A verified, step-by-step OpenClaw 2.0 setup guide for beginners: the Node version you need, the three ways to give the agent a model and what each costs, the exact six-step guided onboarding introduced in v2026.8.1, the four-step Telegram pairing flow, three first tasks worth running, the day-one permission settings that keep the blast radius small, and the five failures new installs actually hit.
On August 31, 2026 an open-source AI agent released v2026.8.1: 16,000+ pull requests from 933 contributors, roughly half of every PR the project has ever merged, after 49 days of deliberate silence following 106 releases in 230 days. The setup wizard now imports the ChatGPT or Claude subscription you already pay for, or finds your local Ollama models. Sessions moved to SQLite and downgrades are no longer free. A close read of what 'AI you own' actually looks like — including the breaking changes and the defaults that are still off.
Bangkok agencies publish consecutive interpreting from 3,000 THB an hour and simultaneous from 5,000, and consecutive roughly doubles how long the meeting runs. Google Translate's Live translate covers 74 languages including Thai, Google Meet's translated captions cover 69 including Thai — and Meet's Transcripts and Gemini note-taker cover 8 each, neither of which is Thai. Deepgram's Nova-3 transcribes Thai as th and th-TH but excludes it from the 10-language code-switching mode, which is exactly the Thai–English sentence your meeting is made of. A full accounting of meeting translation options for Thai teams.
Notta's free plan grants 120 transcription minutes a month and caps a single recording at 3 minutes. TurboScribe gives 3 files a day with a 30-minute ceiling. Google's API hands you 60 free minutes and no interface. On August 7, 2026 Deepgram shipped an improved Indonesian model — on the batch path only. A tested comparison of the free tiers a journalist, researcher, or HR interviewer in Indonesia actually runs into, and the limits none of them advertise.
Word error rate scores the word 'not' and the word 'the' identically, which is why two transcripts can both land on 10.0% and only one of them be usable. Here is the full measurement procedure: a 12-minute test clip built to meeting specs, a reference transcript with frozen normalisation rules, the exact cpWER and DER commands, and a five-dimension scorecard that gives raw WER 10 points out of 100.
DeepL launched Voice-to-Voice on April 16, 2026 across 40+ languages, and the same product's FAQ says meeting data is 'processed temporarily in memory and deleted once the call ends.' On August 28 the Open ASR Leaderboard added its first Global South language and showed two models tied at 4.9 WER differing almost fourfold in how much their accuracy depends on where the speaker is from. Real-time meeting translation is becoming a commodity. The record is not.
On August 20, 2026 a model with no lab, no card and no price appeared on OpenRouter as stealth/ox-alpha. Four days later OpenCode users had pushed 26 trillion tokens through it. Six days later Z.AI put its name on it: GLM-5.3-Flash, 320B total parameters, MIT weights, $0.15 per million in. Why labs now ship models without their own name attached, what the traffic numbers really measure, and what a 24-cent blended price does to the economics of speech pipelines.
On August 4, 2026 Deepgram added Punjabi and Nepali to Nova-3; on August 7 it added Armenian and improved Tamil, Indonesian, and Belarusian. Nine months ago the model covered 31 languages — today the docs list 58. But code-switching still works in exactly 10, and August's own changelog shows Indonesian improving in batch while Belarusian improved only in streaming. A look at the pace of speech-recognition language coverage, and at what the word 'supported' is actually hiding.
Google Meet's translated captions cover 69 languages including Vietnamese, and Google Translate's Live translate covers 74. Meet's Transcripts feature and its Gemini note-taker each support exactly 8 languages, and Vietnamese is in neither. DeepL Voice can only transcribe spoken Vietnamese through a third-party provider that has to be enabled by an admin and cannot use auto-detect. A look at what live voice translation actually gives Vietnamese teams in 2026 — and what it never gives them.
Three years ago the best model you could download lost to GPT-4 by 37 points on HumanEval and needed two graphics cards to run. Last week a 27B open-weight model scored 52.0 on Artificial Analysis's Intelligence Index against Claude Opus 4.6 Max's 44.9 — at one-ninth the price, from a 16.5 GB file. A before-and-after with the benchmark tables, the price curve, and what it means that the catch-up lag is down to 162 days.
A product walkthrough of the full Telli.sh pipeline — live recording or audio upload, AI meeting transcription with speaker identification, real-time meeting translation across 44 target languages, AI summaries in 10 formats, and a shareable note. With the honest limits, the exact prices, and what meeting transcription accuracy really requires.
Research on meeting-transcript summarization found that 30.4% of generated summaries of city-council meetings carried at least one factual error, and seven distinct error types behind them. A practical guide to reviewing an AI meeting note against its own transcript: what to check, in what order, and what to say in the room so the review takes nine minutes instead of thirty.
AirPods Live Translation runs 10 languages entirely on your iPhone. Timekettle's $449 W4 Pro covers 52. Pixel Buds do 40 but need a connection. We read all four vendors' own support pages on 18 August 2026 to answer a question the marketing skips: when the conversation is over, which of them still has it? A guide to the translator-earbud landscape, and to the second job nobody sells you.
On August 11, 2026, Anthropic began marking Claude's written output so it can later be traced back. The mark survives copy-paste and even translation, and disappears under a rewrite. A plain-language guide to how text watermarking works, what a detection result does and does not prove, why the false-accusation history matters, and why every one of the 1,922 hallucination cases now on record would have passed the check.
Qwen3.8-27B ships under Apache 2.0. The Q4_K_M GGUF is 17.1 GB and fits a 32 GB laptop; FP8 is 30.9 GB. Measured: 48 tokens/s on an RTX 4090, 12.75 on a Mac mini M4 Pro.
Alibaba's Qwen3-ASR models are Apache-2.0, cover 52 languages and dialects, and download in under two gigabytes. A plain-language guide to the open-weight transcription layer: what exists, what running a model on your own laptop actually buys you, and where the closed APIs still win.
Telli.sh paid plans came down again on July 20: Pro is $8 a month, Pro+ is $10, and yearly plans keep the 15% discount. Here is the old-versus-new table and why we rounded the numbers down.
In three weeks China's biggest AI labs shipped or teased frontier open-weight models — Kimi K3, Qwen3.8-Max, and DeepSeek V4 at general availability. A plain-language guide to what each one is, what actually shipped versus what is still a promise, and why Beijing may now restrict the wave it enabled.
OpenAI's July 2026 research found that 43.5% of occupation-specific ChatGPT messages are about work outside the user's own job. That kind of work runs on context — and most of an organization's context evaporates in meetings. A practical guide to turning meetings into searchable work knowledge: what to capture, how to name it, which meetings deserve a durable record, and what search changes for onboarding and handovers.
The bottleneck isn't the meeting — it's the day after, when decisions lose their edges, owners blur, and follow-ups die in DM threads. A practical walkthrough of the note-taking loop that execution-first teams run: capture the decision, owner, and date live; distribute within minutes; reopen the note at the start of the next meeting.
Microsoft's MAI-Transcribe-1.5 posts a 2.4% word error rate — and its own documentation says diarization isn't supported. A plain-language guide to what WER measures, why the leaderboard has bunched up near the floor, and which layers above raw speech recognition actually decide whether a transcript becomes a usable meeting note.
Granola put an AI notepad on your wrist, VOMO passed 400,000 users, and open-source clones arrived in weeks. None of it changed what makes a meeting note useful. Here are the six things that still have to be in one, and a 60-second test for your last note.
Kimi K3 is the headline of the 2026 open-weight surge, but GLM-5.2 is the model you can actually download today. Here is an honest read on both, plus why permissive open models matter for meeting notes.
OpenAI routes the gpt-5.6 alias to Sol, with Terra for balance and Luna for volume. Reasoning effort now runs none to max, default medium — what that changes when you migrate off GPT-5.5.
Anthropic's Fable 5 access pause shows why safer AI models still need clear safeguards, source records, and explainable user-facing controls.
Deepgram's diarization v2 was preferred 3.3x over the previous production diarizer — but diarize_model is batch-only, not streaming. Plus Korean spacing and multilingual numerals.
A dated snapshot: on June 28, 2026, OpenAI's latest-model guide still pointed to GPT-5.5, not GPT-5.6. What the guidance claimed about reasoning, tool use, and long-context execution.
We lowered Telli.sh paid plan pricing so more people can use live AI notes, transcripts, summaries, and multilingual meeting workflows.
Google is pushing live speech translation into Meet and Translate. Telli.sh focuses on the part live translation can miss: a reviewable record of what was said.
South Australia Police is preparing an AI speech-to-text trial for interviews. The lesson for every serious meeting workflow is clear: speed only works when the record stays reviewable.
Telli.sh: Every Translator has launched on the Chrome Web Store with free Google Translate mode, Chrome built-in on-device AI, and translation beside the original text.
Telli.sh can now start from a live recording or uploaded audio and turn conversations, lectures, interviews, and presentation reviews into reviewable AI notes.
A practical first-note workflow for teams in the Philippines: upload one meeting recording, check the transcript, then use the summary and follow-ups after reviewing the source.
Telli.sh now recovers Deepgram live transcription streams when the upstream connection drops so Pro and Pro+ recordings can continue without forcing a browser restart.
We turned Sentry back on, fixed the bugs it surfaced, and tuned the integration to live comfortably inside the free tier. Here is what the data told us — and what the observability market looks like in 2026.
AI notetakers are moving beyond transcription into meeting intelligence, real-time context, and trusted follow-through. Here is how the latest audio AI shift shapes Telli.sh's direction.
A simple look at the recent fixes that made recording and deployment feel more reliable.
Meet Telli.sh — real-time transcription, translation, and AI meeting notes that break language barriers.