Jev by @typesafeai is about to flip agent architecture on its head.
You're about to end up with 4 steps in your voice pipeline, with the LLM being replaced by a thinker (Jev / RLCD) and a talker.
This is going to reduce costs dramatically, and potentially make agent adoption faster
lukeocodes.dev/jev-system-on…
Over the last 24 hours we've released signifcant efficiency improvements to Simba 3.2, our flagship TTS API model, and Simba 3.0 our current generation multi-lingual model
If you've been paying attention, you'll already know how well we perform in human blind testing, how accurate we are, and how easily you can get up and running.
We're still just getting started!
speechify.ai/build
(Screenshot from benchmarks.coval.ai/tts?wind…) @SpeechifyAI
OpenAI, Anthropic and Google DeepMind have reportedly met since July about an AI standards body, two months before the essay calling for restraint...
lukeocodes.dev/ai-standards-…
We just raised $4M at @lightsagehq , led by @nexusvp , to build the growth stack for internet’s newest customer: AI agents.
Claude Code, Codex, Cursor and other agents can already discover products, compare alternatives, read documentation, install SDKs, call APIs and start using software on behalf of humans. Soon, they’ll increasingly transact too and that changes how internet companies grow.
For decades, companies built growth infrastructure around humans. PLG helped humans adopt products. Now there’s a new journey to optimize: Can an AI agent discover you, choose you, successfully use you, and eventually pay for you?
We call this Agent-Led Growth.
Lightsage helps companies get discovered, recommended and used by agents
→ Agent Discovery - when and why agents recommend you vs. competitors
→ Agent Experience - whether agents can successfully use your APIs, SDKs, CLIs, MCP servers and agent skills
→ Agent Analytics - what real agents do when they reach your website, docs and product. and most recently agent-attributed traffic conversion.
Companies including @firecrawl , @reductoai , @daytonaio , @rimelabs , @Tiny_Fish and @insforge are already using Lightsage to build their agent-led growth strategy and improve their agent journey.
We’re incredibly grateful to have Nexus lead this round alongside these amazing founders and operators.
- Former Salesforce CTO Steven Tamm
- Postman CEO @a85
- Apollo CEO Matt Curl
- DocuSign President Robert Chatwani
- Former GitLab Head of Growth @HilaQu
- Resend CEO @zenorocha
- Firecrawl Co-founder @ericciarla
- Daytona CEO @ivanburazin
- Tinyfish COO @shuhao_friday
- Adam Frankl
and many more!
As agents increasingly discover, evaluate, use and pay for products on our behalf, we believe every company will need to understand and optimize for the agent channel.
That’s what we’re building Lightsage for.
Huge thank you to my co-founder Sean Er, customers, early investors @AlchemistAcc and everyone who believed in us early. Honoured to be featured in @Forbes .
We’re just getting started.
lightsage.com
Simba 3.2 by @SpeechifyAI is the best value TTS model right now.
It ranks #2 overall in our human preference benchmark while costing just $10/1M characters, putting it in a league of its own.
So he asked the chatbot what went wrong. It explained that its previous advice was terrible for his crops. It diagnosed its own crime with the detachment of tech support closing a ticket.
Friday fun: a farmer asked a chatbot how to protect his sesame crop. It prescribed a broadleaf herbicide. Sesame is a broadleaf species. 100,000 square meters gone.
Practical takeaway for voice AI engineers: plan for context tracking from day one. Even if your current system is turn-based, design the state layer to accept playback position callbacks.
Adding it later is a rewrite.
lukeocodes.dev/full-duplex-v…
The research community is responding:
• ReLMCodec — predictable speech tokens so LLMs can plan prosody
• CtrlSpeech — coarse-to-fine control for mid-flight adjustment
• Beyond Naturalness — we're evaluating the wrong thing
Each tackles a different piece of the puzzle.
The relay race architecture (STT→LLM→TTS) breaks in 3 ways in full-duplex:
1. Pipeline latency accumulates — every ms pushes response further behind
2. State is shared but playback isn't — model can't tell what user actually heard
3. Interruptions are hacks, not features — barg…
The core problem: LLM-based voice systems advance dialogue state faster than audio plays. So the model makes decisions based on content the user never heard.
You interrupt to clarify. The model already moved on.
Full-duplex voice AI has a context problem nobody's talking about.
The model generates responses based on audio the user never heard. PACE calls it Generative Context Mis-anchoring.
The relay race architecture can't fix it. Here's what needs to change.
lukeocodes.dev/f…
TIL: GitHub Actions $/ syntax — reference same-repo actions at the exact running commit. No more hardcoded versions or ./ path hacks. lukeocodes.dev/til-github-ac…