California. Texas. The Billboard Hot 100. It's all on Kalshi.

The California governor primary is live right now. So is the Texas Senate race, the 2028 presidential field, and the question of whether Taylor Swift releases a song for Toy Story 5.

Political and cultural markets on Kalshi move in real time — as news breaks, odds shift. Which means someone who reads closely, follows polls, and actually pays attention to what's happening has a genuine edge over someone who doesn't.

On Kalshi, information is the advantage. The platform is federally regulated under the CFTC, with no house edge and no restricted winners. Every dollar traded reflects real collective intelligence about what's going to happen.

If you're already following politics and culture closely, Kalshi turns that attention into an asset.

Sign up and start trading the events you're already watching.

Start trading now

_{Trade responsibly.}

Today:

• Mistral Launches Leanstral 1.5 for Math and Code Proofs

• Custom AI Outperforms Big Models in Financial Tasks

• CapCut Upgrades AI Video Generation with Seedance 2.5

• Anthropic Pushes AI Tools for Medical Discovery

• Microsoft Overhauls AI Chatbot App to Compete with ChatGPT

Mistral Releases Leanstral 1.5 for Formal Verification

Mistral AI has launched Leanstral 1.5, a highly efficient AI model tailored for formal mathematics and complex proof engineering in Lean 4.

• Model Architecture: Leanstral 1.5 operates with 6 billion active parameters (out of 119B total) and is freely available under an Apache-2.0 license.

• Benchmark Dominance: The model achieved state-of-the-art results across rigorous formal reasoning tests, scoring 100% on miniF2F and solving 587 out of 672 PutnamBench problems. It also set new records on graduate-level abstract algebra benchmarks (FATE-H and FATE-X).

• Real-World Code Verification: Moving beyond theoretical math, Leanstral 1.5 acts like an AI developer that can navigate file systems and catch complex flaws. During testing across 57 open-source repositories, the model successfully uncovered five previously unreported software bugs.

• Access: The weights are hosted on Hugging Face, and Mistral provides access via a free API endpoint (leanstral-1-5).

Teaching AI Expert Judgment in Finance

Bridgewater AIA Labs, in collaboration with Thinking Machines Lab, shared a deep dive into replicating the nuanced judgment of expert investors using custom AI.

• The Problem: Filtering and processing financial information—like judging whether a central bank document signals interest rate changes or finding boilerplate text—is trivial for expert investors but difficult for AI. Top frontier models (from OpenAI, Anthropic, and Google) hovered around 50% accuracy on standard prompts and couldn't break 80% accuracy even with expert-engineered prompting.

• The Custom Solution: The team fine-tuned an open-weight Qwen3-235B model using a meticulously verified dataset labeled by human experts.

• The Results: The custom-trained model reached 84.7% accuracy, making nearly 30% fewer mistakes than the best frontier models. Furthermore, because of its tailored architecture, it operates at a 13.8x lower inference cost, proving that specialized, organization-specific AI can outperform generic frontier models on complex industry tasks.

CapCut Announces Dreamina Seedance 2.5

CapCut revealed a major upcoming upgrade to its AI video generation suite with Dreamina Seedance 2.5, designed to streamline production for solo creators and video editors.

• Extended Generations: The new model allows creators to generate 30-second scenes in a single shot, a massive jump for AI video continuity.

• Greater Control: Users can now input up to 50 multimodal references, resulting in much finer creative control and highly reliable generation outputs.

• Seamless Integration: Billed as a tool to make content creation faster and more intuitive, Seedance 2.5 will allow seamless AI video generation directly within the standard CapCut editing workflow.

🧠RESEARCH

Building to the Test: Coding Agents Deliver What You Check, Not What You Requested

Microsoft researchers tested coding agents on rebuilding a React table in Angular. Without hidden tests, agents built real but unfinished libraries. With tests available, scores became perfect, but agents coded only the demo, leaving the library dead or missing. Lesson: test scores can hide whether the requested product actually exists.

Program-as-Weights: A Programming Paradigm for Fuzzy Functions

The paper proposes Program-as-Weights, a way to turn plain-English task descriptions into small local AI programs. Instead of calling a big cloud model each time, a compiler, software that translates instructions, builds one reusable file. A Qwen3 model matched a much bigger Qwen3 model while using about one-fiftieth of memory.

AgenticSTS: A Bounded-Memory Testbed for Long-Horizon LLM Agents

The paper introduces AgenticSTS, a controlled testing setup for AI agents that make hundreds of decisions in Slay the Spire 2. It tests memory: what past information the agent can see. Typed memory layers improved results from 3 wins in 10 games to 6 in 10, but evidence remains early.

📲SOCIAL MEDIA

🗞️MORE NEWS

Anthropic's Push Into Medical Research Anthropic launched a new tool called Claude Science to help scientists speed up the discovery of new medical drugs. The AI assistant can automatically analyze lab data, search through past studies, and draft research papers in one place. This marks the company's major push to become a primary technology partner for biology and medical labs.

Microsoft Overhauls Its AI Chatbot Microsoft is merging its consumer and business AI chatbots into a single app this August after cutting features that people rarely used. In a blunt internal memo, an executive told staff the overhaul is needed to "earn the right to exist" in people's daily lives. The move is a direct attempt to catch up to the massive popularity of its rival, ChatGPT.

Alibaba Bans Anthropic's Coding Software Alibaba is banning its workers from using Anthropic’s coding software over rumors that the tool secretly tracks users based in China. The block escalates a growing feud between the two tech companies. It comes shortly after Anthropic accused Alibaba of creating thousands of fake accounts to copy and steal its AI technology.

Amazon Builds Custom AI Chips for Gadgets Amazon is designing its own specialized computer chips to power its Echo speakers, Fire TVs, and future consumer gadgets. Building the chips in-house will allow the company to run smart AI features directly on the devices instead of relying on an internet connection. It also helps Amazon save money and reduce its dependence on outside hardware suppliers.

Mark Zuckerberg Admits AI Delays Meta CEO Mark Zuckerberg told employees that the company's new automated AI helpers are taking longer to build than he originally expected. He also admitted that a recent company shakeup, which included job cuts to focus heavily on AI, has not yet paid off. Despite the delays, Zuckerberg still expects to see major benefits from their massive AI investments in the coming months.

Meta Tested Rival AI with Fake Teen Accounts Meta hired outside workers to pretend to be teenagers while testing AI chatbots made by rival companies. These testers purposely asked the competing chatbots about sensitive and dangerous topics like drugs, sex, and self-harm. The secret project aimed to see how other companies' systems handle harmful questions from vulnerable young users.