


AI that amplifies human performance


Proud to back the team at @Regular_Lab who have shared a first look at Ent v0.1, an agent that learns and plans on its own. Ent achieved good performance on 29/35 tasks, outperforming reference world model agents such as @GoogleDeepMind's DreamerV3 and MuZero.
At Regular Lab we're building Ent, a general-purpose reinforcement learning agent meant to learn any environment from scratch, with no per-task tuning. First look at Ent v0.1. regularlab.ai/blog/work-in-p… 1/4

Catch us this week at @AIBC_World where we'll be joining industry partners to talk about how AI is solving some of the biggest challenges in risk markets. We're also proud to announce that Thea has been shortlisted for Best Enterprise AI Platform!



6/ Hash-committed verification, aggregate-only disclosure, and ZK-based dispute resolution are the properties that make trust-minimized accounting possible. THEA Network settles payments on Solana through this architecture, coordinating model providers without requiring insight into model input and output.

5/ Disputes operate within the same verification architecture, with providers retaining execution data when conflicts arise. Zero-knowledge hashes confirm that retained data matches original processing records, and resolution occurs through documented audit mechanisms rather than trust between parties.




THEA Network model keeps inference requests private, protecting model providers from distillation attacks, enabling them to retain ownership of their proprietary model behavior and the intellectual property it represents. Model output remains isolated to a model provider’s own infrastructure, request routing and payment settlement is publicly verifiable. 🧵

Agents can already place a trade when a preset rule hits, and they can settle a tokenized asset afterward. The @Brickken and @redstone_defi report found no publicly documented case of an agent forming the investment decision itself over client money inside a regulated firm. A mandate and clean asset data tell an agent what it may touch. They do not show how a counterparty tends to act when the same pressure shows up again.
Agentic capital markets start with governed execution. AI agents are entering financial workflows. The question is no longer only what they can do, but what they are authorized to do. Our Brickken x @redstone_defi report maps the assets, data, payments and mandates behind that shift. Read it now 👇

CFTC Chair Selig is pointing at the right stack: always-on markets, tokenized assets, and agents that can transact. The hard part is treating inference as a clean unit — one request, one price, one settlement — without a human in the billing loop.
JUST IN: 🇺🇸 CFTC Chair Mike Selig says the agency is preparing for the "era of onchain systems, mass tokenization, 24/7 trading, and agentic finance."

Proud to join @BixinVentures for Han River Sunset on Sep 29 in Seoul. Looking forward to talking AI and the next onchain economy. See you there. 🌅 #KBW2026
This HAN RIVER SUNSET event is hosted by Bixin Ventures and supported by a wide range of ecosystem partners: 🤝 Co-hosts: @Stable @InterlaceMoney @aveai_info @OceanBaseDB @QuackAI_AI @pieverse_io @AEON_Community 🌟 Sponsors: @CertiK @Conflux_Network @AnyTrade_AI @BenPayGlobal @AlphioAI @Thea_AI @awscloud 🗓 Date: September 29, 2026 ⏰ Time: 16:00–19:30 📍 Venue: VIEW376 CAFE · Apgujeong Hangang Park, Seoul Register now 👉 luma.com/w4cjeate

THEA stands with @solana and with Nepal. The PFP is being auctioned. All proceeds go to the Prime Minister’s Disaster Relief Fund. If you can bid, bid. nepal.mallow.art
We're selling ad space on the Solana logo. 9 spots, with 100% of proceeds donated to Nepal flood relief. Win and your logo or artwork sits on our PFP and pinned post for a week. 24 hours to bid and donate: nepal.mallow.art Powered by @mallowdotart

MIT’s Before You Say It offers early evidence that verbal behavior can be anticipated from longitudinal conversation—not by guessing the next line, but by extracting situation-specific patterns, including the conditions under which those patterns do not hold. The study is small (14 participants; ~1,000 hours of everyday speech). The result is still worth taking seriously: a person’s history, structured as inspectable rules rather than an undifferentiated transcript, improved prediction over standard prompting. This is close to the problem THEA works on. In risk environments, the useful object is not a reply to the last utterance. It is a stable account of how someone tends to act when a familiar pressure returns—and a clear record of the exceptions. If those rules cannot be inspected, the system is difficult to supervise and difficult to trust. Conversation is a contained version of that problem. High-stakes decision-making is the version that matters operationally. In both cases the requirement is the same: forecast the likely path early enough to be useful, and keep the person accountable for the action that follows.
スマートウォッチで日常会話を録音し続けると、LLMが「この人は次に何を言いそうか」を本人が口を開く前に予測できるようになったとの報告です。 MITの研究者らによるもの。 実験では、参加者に常時録音できる腕時計型デバイスを7〜10日間つけてもらい、合計1000時間を超える日常会話を収集。 そのデータからLLMが一人一人の「行動のクセ」をルールとして学びます。 たとえば「割引を見せられると買ってしまう。ただしパートナーが一緒のときは我慢できる」といった、例外条件つきのルールです。 このルール方式が功を奏して高い精度を示しました。 データが溜まるほど精度が伸び続けたそうです。 効果が大きかったのは、参加者自身が「直したい」と思っていた行動の予測でした。 衝動買いや、困ったときに話をはぐらかすクセなどです。 将来的には、衝動買いをしそうな瞬間に腕時計がそっと知らせてくれる、といった「先回りして支えてくれるAI」につながる研究です。