SoundHound swallows LivePerson for pennies, calls it OASYS

Good morning 👋 SoundHound just bought a Nasdaq-listed rival for less than a seed round. The bargain price tells you more about the conversational-AI market than the "powerhouse" press release does.

In today's issue:

• 🔭 SoundHound swallows LivePerson for pennies, calls it OASYS
• 🧠 NVIDIA/AWS ship a physical-AI training pipeline; AWS open-sources HyperPod's control plane
• 🔬 Shorter prompts that work better, a repeat-yourself instruction-following trick, and agents that over-edit code
• 🛠️ Auditing your channel stack before your vendors consolidate for you
• 🔗 Intuit's agentic disaster-recovery assistant, and an AWS team's dashboard-failure detector

🔭 THE ONE THING

🎙️ SoundHound swallows LivePerson for pennies, calls it OASYS

SoundHound AI closed its acquisition of LivePerson on September 4, five months after signing, folding the two into a platform it's branding OASYS: 25 Fortune 100 customers, 750+ patents claimed, and John Collins, LivePerson's old CFO, now running finance for the combined company. The price tells the real story. $43 million in equity, 0.47 SOUN shares per LPSN share, is a rounding error for a company that spent over a decade as the public face of enterprise chatbots before LLM-native rivals ate its lunch. SoundHound also retires LivePerson's debt and pockets roughly $74 million of its cash, so this reads less like a platform coup and more like a distressed-asset pickup with a rebrand stapled on. Watch what it does to SoundHound's own balance sheet before buying the "powerhouse" framing.

🧠 MODELS & RELEASES

• 🏭 NVIDIA / AWS shipped a reference build for running NVIDIA Cosmos 3's synthetic-data-to-eval loop on a persistent SageMaker HyperPod cluster. Physical AI (robots, AVs) needs continuous data generation, post-training, and closed-loop eval, not a one-off run, and this is AWS packaging that as infrastructure. Build a Physical AI model factory with NVIDIA Cosmos 3 on SageMaker HyperPod
• AWS open-sourced HyperPod InstantStart, a control plane that runs cluster bootstrap, capacity, training, and inference through either a UI or an AI agent, same guardrails either way. Worth a look if you're tired of hand-holding SageMaker clusters through YAML. Run agent-driven SageMaker HyperPod operations with InstantStart

🔬 RESEARCH HIGHLIGHTS

• ESPO fixes prompt bloat. GEPA-style evolutionary prompt optimizers keep bolting on caveats until prompts triple in length without getting any more accurate. ESPO diagnoses the actual error clusters first, proposes targeted fixes, then bootstrap-tests them before keeping one. Result: prompts 47% shorter (1,004 vs. 1,878 characters) and 3.76 points more accurate than GEPA across seven benchmarks. On Qwen3's GSM8K split the gap was absurd: 15% to 91%. Builder read: if your prompt-tuning loop just piles on rules when something fails, you're doing the thing this paper fixes. Paper
• Say it twice, get it followed. Repeating a procedural instruction partway through the prompt, no fine-tuning, no architecture change, cut instruction-following failures across seven models and 16,800 test runs. Final-answer accuracy barely moved, so this is about compliance with the process, not the outcome. One tradeoff: models jumped to conclusions slightly more often. Cheap enough to try before reaching for anything heavier. Paper
• Your coding agent rewrites more than it needs to. Researchers corrupted 400 BigCodeBench solutions and measured how much extra code models touched while fixing them, calling it "over-editing." Even GPT-5.5 does it: a high pass rate hides unnecessarily large diffs. Just instructing the model to preserve existing code cut excess edits and nudged pass rate up too. RL training taught minimal edits better than supervised fine-tuning, which just overfit to patterns it had seen. Watch, don't act: worth a system-prompt tweak on your review agent, not a pipeline rebuild, until this holds up outside one benchmark. Paper

🛠️ TRY THIS

Audit your channel stack before your vendors consolidate for you

SoundHound just closed LivePerson, folding two separate conversational AI vendors into one omnichannel platform. If you route customer conversations through either, your integration surface is about to move whether you asked for it or not.

1. List every customer-facing channel you run today (voice, web chat, SMS, WhatsApp) and which vendor's API sits behind each.

2. Flag any channel currently split across two vendors mid-acquisition. That's the seam your account rep hasn't briefed you on yet.

3. Get a written migration timeline and a price lock in before your next renewal, not after the merged roadmap ships.

4. Prototype a fallback: could you fail one channel over to a different vendor inside a week if the combined platform slips?

Prompt: List every customer-support channel we run (voice, chat, SMS, in-app) and the vendor behind each. Flag any that map to a company currently mid-acquisition, and draft three questions to ask that vendor about migration timeline and price lock before our next renewal.

Worth a look

• Hugging Face shipped a full TRL + OpenEnv recipe for RL-training a 1.7B coding model to paint with p5.brush: GRPO, LoRA across the MoE's linear layers, a reward split between a vision judge and an aesthetic scorer, on a single A100. Hugging Face

🔗 QUICK LINKS

• Intuit built an agentic disaster-recovery assistant on Bedrock, EWOK Agent, that runs production failovers from a plain-language request and logs every action for audit. Nice template if your on-call runbook still lives in someone's head. AWS blog
• AWS open-sourced how one internal team catches BI dashboards that go quiet or wrong while every infra check still says green. If your data team trusts a dashboard just because uptime looks fine, it probably shouldn't. AWS blog

See you tomorrow.

Pradeep Perugu