Potentially biased source.. but I can believe this faux productivity.
Volume is up, quality is down. Were offloading our critical thinking, long term outcomes totally unknown.
Also, bit of an overreaction no?
If you are smart and good at computers you should get the fuck out of this country the moment you can - it is over - go to America and earn 3x what you’d get here then come back rich. We are Poland with a tan
@TravelodgeUK Hi, I've asked to move a booking at London Central Waterloo (20-22 Oct) and my case reply time is 5-8 days, after check-in on Tuesday. Could someone look at it before then? Happy to DM the details.
These are cool but using Baileys risks WhatsApp terminating your account… which sucks. I wish WhatsApp had a plan to open up access for agents like this.
we one-shotted an open-source Instinct.
meet inkling: an AI that lives in your WhatsApp gc. it runs the poll and adds it to everyone's calendar. reminders, voice notes and memes too.
see it, say it, self-host it.
built with @ask_joon
We’re working on making apps and agents work underground, in plant rooms, lift shafts, metro stations, etc.
What are the most popular offline first apps, other than @linear? Looking for good detailed write ups on real world experiences building these.
Instinct in group chats
Starting today, early access users can add Instinct to group chats. A new Instinct joins and works for the whole group.
Making plans with friends usually turns into a frustrating back-and-forth over times and places. With Instinct in the group, you can explore options together, agree on a plan and get it done, all in one thread. Your friends don't even need Instinct to join in.
A few things it’s good at:
- Planning a weekend trip with friends, including dates and arrival times
- Coordinating logistics with a roommate
- Getting tickets as soon as they go on sale, then splitting the cost and sending one to everyone
- Running a fantasy league, keeping standings, settling rules disputes, and reminding everyone to set their lineups
- Sorting out carpool schedules - works out who’s driving each day from everyone’s schedules
- Organizing Thanksgiving, keeping track of who’s bringing what and ordering the shopping list
Group chats are new territory for Instinct, and we put a lot of care into getting the permissions right. Here’s how it works:
- Your personal Instinct asks before connecting with the group’s Instinct. You choose which groups to trust and can remove trust at any time.
- The group’s Instinct has no direct access to your personal accounts. Your Instinct checks your permissions before sharing information or taking action.
- If someone new joins the group, all pending replies from your Instinct are held until you grant it permission to share with the expanded group.
Group chats are available to all of our early access users, and will be rolling out to everyone soon. To join early access for this release, ask your Instinct to put you on the list.
Great work with @GajanNagaraj@anuda_w and @saagarwashere
@lawrjones Ok nice - very similar to us, except we don't store the fixtures in git. It's in GCS.
We're also experimenting with customer specific evals given we allow them to describe behaviour in natural language. How these evolve, and how we manage them, is challenging to say the least.
@lawrjones Right, and for fixtures that are too.. large? All in the YAML, such as tool calls, or user policies? or are most of them single turn input/output style tests?
We have both evals, but also more classic ML datasets like transcription, classification, etc. They all feel the same.
@evnsio I'm going to counter that min/max take with my own.
Not looking, let alone understanding the code your agents produce is just accepting you're happy with high interest tech debt.
make the claude slop stop. wtf does this mean?
"The P2 splits into a valid kernel and an invalid majority — let me verify the from-record leg before fixing, since that's the one path cubic lists where the SR is referenced downstream".
Improving a product with a chat interface is so much easier than a website. With a website, you're stuck watching screen recordings and guessing what users wanted. With a chat interface, they literally tell you in English what they want your product to do.
This is the classic min/max tweet that adds nothing to the conversation about voice agents.
The real headline: NVIDIA released an iteration on speech to speech agents. We're still a long way from these being useful in production.
The voices suck, tool calling with be bad, etc
The part most people will skip: NVIDIA just made every voice AI API a commodity.
OpenAI charges $0.06/min input and $0.24/min output for Realtime API. Gemini Live bills 25 tokens/second of audio. Every startup building voice agents is hemorrhaging cash on per-minute API fees to run what is fundamentally a pipeline problem: ASR → LLM → TTS, three models stitched together with latency at every seam.
PersonaPlex replaces that entire pipeline with one 7B model. Runs on a single A100. Open weights, MIT license, commercial use permitted. Response latency: 0.170 seconds for turn-taking, 0.240 seconds for interruptions.
It scores higher on dialog naturalness than Gemini (2.95 vs 2.80 MOS) and handles interruptions better than every commercial system they benchmarked.
This tells you everything about NVIDIA’s playbook. They don’t need to charge for the model. They need you to buy the GPU. Every company that self-hosts PersonaPlex instead of paying OpenAI per-minute is another A100/H100 sale. Every voice agent startup that drops their API dependency is another enterprise GPU contract.
NVIDIA open-sourced the fishing rod because they sell the lake. Built on the Moshi architecture from Kyutai, fine-tuned with under 5,000 hours of data.
The voice AI margin is migrating from the application layer to the hardware layer. And NVIDIA is the only company that profits no matter which model wins.
330,000 downloads in the first month. That’s infrastructure capture disguised as generosity.
At @Elyos_AI We benchmarked 13 STT providers on 100 real customer calls from the trades businesses.
Not synthetic lab data. Real calls with:
- Background noise & multiple speakers
- UK postcodes & addresses
- Regional accents (England, Scotland, Ireland)
- Short confirmations to long explanations
Top performers:
🥇 @DeepgramAI Flux - 15.9% WER
🥈 @soniox_ai - 16.9% WER
🥉 @Speechmatics - 17.7% WER
@OpenAI Whisper? 39.8% WER - wouldn't recommend for production voice AI.
What's your experience with STT models? Are we there yet?