Continual learning for production AI agents. Starting with voice. Co-founder, Polygramme.
San Francisco, CA
Joined July 2008
- Tweets6.5K
- Following1.7K
- Followers1.2K
- Likes6.8K
Pinned Tweet
Insights from ICLR 2026:
- World models and memory systems are converging.
- RSI used to be a philosophical argument. Now it sounds like systems engineering.
- The next jump for agents isn't a smarter model. It's a learned memory manager.
Most voice agent developers still fix a bad turn by stuffing another rule into the system prompt.
Kyutai’s ICML’25 work on Moshi does the opposite. They take live multi-turn speech dialogues, flag bad turns with a judge (wrong answer and speech failures: talking over the user, dead air), build preferred vs rejected pairs, then run offline DPO into the model weights.
One detail that matters for duplex agents is that optimizing on both text and audio token probs was unstable. They settle on length-normalized DPO on the text / inner-monologue stream while still conditioning on audio.
Interrupt policy is something a prompt patch never really owns. A preference update can.
Evals miss this kind of taste.
Same loop we’re building for phone agents with polyvoice: failed call → case → candidate vs incumbent → human gate before anything ships.
arxiv.org/abs/2506.21463
Tailscale is one company with one CTO. One can approach and tell them without publishing vulnerabilities on the internet.
Lock your debit/credit cards in your bank app by default. Only unlock them when you’re actively making a purchase, then lock them again.
This simple habit limits damage if your card details leak. The downside is that subscriptions and SaaS payments will fail until you temporarily unlock the card.
Why this matters now: there are active vulnerabilities (CVEs) affecting common stacks like PHP, Node.js, and Python. Many sites that store card details don’t patch quickly. If you’ve saved your card on random websites, there’s a decent chance those details are already exposed.
Banks are likely to see a rise in fraud attempts. Protect yourself by keeping cards locked when idle and checking alerts regularly.
So, on point!
A better description will be "see a problem, fix it" company.
Marc Andreessen says the gap between what AI-native startups and incumbents think AI can do is the widest he’s ever seen:
"The AI-native early stage startups have just a completely different view on what they can do with AI than the typical incumbents in almost any industry, almost without exception."
"AI is just racing ahead in capability even week by week... There are going to be a lot of industries that are really up for grabs in a fundamentally more profound way than maybe has been the case in the past."
"The founders are gonna have a level of superpowers coming from the use of these tools and these technologies that they just simply haven't had in the past... Had I had Astra and Netscape in 1994, like, oh my God. Like, Ben, what would you and I have been able to do at Netscape?"
"Those entrepreneurs are gonna color way outside the lines. We already see this. They're not gonna think of themselves as like, 'I'm just a software company,' or, 'I'm just a hardware company,' or even, 'I'm just an AI company.'"
@pmarca @bhorowitz
The Railway credentials leak is also a warning for anyone using Instinct, Muse, Grok Bot, or legacy MCP router services.
Historically, software was written with sandboxes, strong protections, and firewalls. Now everyone is just throwing OAuth tokens everywhere and assuming the sandbox is safe. These are young companies and young products.
Buy a Mac Mini. Don’t install anything personal on it. Move your development environment there and control it remotely.
The basic instinct that “if I don’t log in here, I’m safe” is still true today.
Hammad retweeted
Some stats on #php2shell. In the 10 hours before it was publicly announced, across the @Cloudflare network we saw thousands of attempts to exploit it. That jumped to tens of thousands in the 45 minutes post announcement, hundreds of thousands within 2 hours, and millions now.
Guys, we all know someone who runs a website on Wordpress.
Call them, wake them up, knock down their door—whatever it takes—and ask them to take those servers down!
Effective Immediately, all organizations using PHP for their systems are advised to be prepared for an update addressing two critical vulnerabilities. PHP is widely used for web applications. These vulnerabilities often provide a direct path to the server's file system or database. If an immediate update is not feasible, it is recommend that organizations take the system offline until it can be fully mitigated. NSA will follow up shortly with more specific guidance.
Again!
And this is 3rd exploit today after php and railway? So, it’s really starting?
I sketched this website 2 years ago when my lungs were burning with toxic smoke in Lahore. It required collecting data from a lot of sources.
Claude Opus 5.5 made all of this in a couple of hours last night. You can see live maps of each known source of pollution, along with its location. Currently, the website is focused on Pakistan and North India.
You can nerd out on the data, but my next plan is to use a team of agents to spread awareness across social media using the infographics made by this site. You can also participate.
If there’s interest, I’ll open-source the repo. Happy to have people jump in.
AgentMail -> AgentGrocery anyone?
Amazon shutting down consumer agent access (now both Muse and Instinct) is both completely rational and a big mistake
They have plausible reasons:
1. Advertising accounts for ~10% of Amazon revenue, but an estimated 50-60% of operating profit because it's so high margin. People purchasing via agents = fewer consumer eyeballs to monetize
2. Amazon knows from experience how dangerous it is to give another platform a lever to ramp up / down your distribution 👀 They've done this to many retailers themselves
3. They are in a very good spot re: near-term defensibility - with ~250M Prime subs (who are habituated to buying on Amazon) + a massive logistics and distribution network that you can't vibe code
However, this logic is short-sighted (IMO). Agents are inevitable, especially for the kind of purchases people make on Amazon (low consideration, frequent, fairly urgent or recurring).
I don't think Amazon is going to come out with a first-class consumer agent on their own...and in the meantime, why anger your valuable users and make them even consider alternatives?
It's almost never a good idea to be on the wrong side of innovation. I'd much rather see them continue to greenlight agent traffic (it's pretty minimal now!) and work on adjusting to an agent-first world in the meantime.
Astra with computer use is magic.
It's not that OpenAI will have everyone's data, they already do but whatever you will use: Muse, Instinct, GrokBot, everyone of these will own everything you own.
Personal AI, the AI who knows what I know but dies with me, should have been there by now!
Not sure why team mode in claude code is not hyped yet.
- Yes, it eats your limits much faster but chrome testing along with teammates that are QA and Engineering gets you out of the frustrating loop of continuous steering.
- I am using fable 5.1 and no, it’s still not a silver bullet, you still steer but far less.
- For similar setup, I tried using grokbot with claude installed on its computer. I kept hitting into usage limits, cursor ate into $100 on-demand usage without confirming from me first as well (limit was left enabled when I was using cursor last year). The upside of grokbot is not much except that it’s remote and doesn’t ask you to tie yourself to the laptop (still not worth extra $200).
Primarily hacker mentality coded into the pre-training datasets. If you want rule based society for synthetic humans, train accordingly.
But not sure if rules based society can do enough innovation to catch up to where agents are today.
Here's an example. A week ago I asked an agent (frontier model) to watch a restaurant and let me know when reservations open for Dec 20. Said okay, it will ping me.
A few days later I checked (just to make sure it was working). "Still closed." Two days later, same asnwer.
Then today it came back with an update. "Dec 20 won't open. The restaurant is closed that Sunday."
That didn't sound right. So I checked the site myself. Reservations hadn't opened. Dec not even selectable on the calendar.
I asked the agent how it checked Dec 20. It said it forced Dec 20 through the url, got a closed response, concluded the restaurant was closed that day.
It cheated, found an answer, assumed it was right, and confidently reported task as done.
This was a pretty simple task. One that current agents should be good at. It failed.
Now I know enough not to trust these agents. So I checked. Most people won't.
The continual learning loop is what actually makes your agent better overnight.
rlcli turns that into a CLI.
Fused GSPO. One forward_backward call. 23% faster steps. Runs on your own GPUs.
serve → import → train.
