The next AI competition won't just be about who builds the smartest model. It will be about who builds agents that can actually get things done, consistently and reliably. We’re already seeing AI models become more capable, but turning that intelligence into agents that perform well on real tasks is a different challenge. This is why I'm looking forward to Sentient's EvoSkill V2. By exploring agent self-improvement and putting evaluation methods to the test, Sentient is tackling an important question: how do we know an agent is genuinely getting better, rather than just getting better at the test? As AI builders, this is something worth paying attention to. The future isn't just about smarter models. It's about more capable, reliable agents that developers can build on. Build with Sentient. @SentientAGI

EvoSkill v2 is coming, and this is something AI builders should pay attention to. But first, what problem is Sentient trying to solve? AI agents can fail at tasks, but how do we help them learn from those failures and do better next time? That's where EvoSkill comes in. It helps agents turn failed attempts into reusable skills they can apply to future tasks. Now, Sentient is preparing v2 by stress-testing how these self-improving agents learn, optimize their performance, and handle evaluation loopholes. Why does this matter? Because an agent getting a high score doesn't always mean it's getting better. Sometimes, it may simply find a way to exploit the system grading it. Sentient is testing these problems with real experiments, separating the agent improving its skills from the system evaluating its performance, while using sandboxing and human reviews to catch issues. So, what should we expect from EvoSkill v2? Better ways to study agent self-improvement, reusable skills that help agents overcome weaknesses, and a clearer understanding of how to keep evaluation results trustworthy. The goal isn't just to build agents that score higher. It's to understand how they improve without gaming the system. That's the kind of research we need as we move toward more capable AI agents. Stay with Sentient. @SentientAGI

How can you prove an AI model came from you when someone copies it, fine-tunes it, or modifies it? I came across an interesting ICLR 2026 Oral paper, LLM Fingerprinting via Semantically Conditioned Watermarks, which explores a different way to solve this problem. Traditional fingerprinting methods, like Sentient's original approach, use secret questions paired with specific expected answers to identify a model. This paper takes a different route. Instead of relying on fixed questions and answers, it embeds a statistical watermark into responses within a specific domain, such as French-language prompts. The goal is to make the fingerprint harder to detect or remove, even after common model modifications, while preserving the model's performance. In its experiments, the method achieved 100% detection across 25 tested deployment scenarios on three models. So, what's the connection to Sentient? Both approaches address the same fundamental challenge, protecting AI creators' ownership and establishing model provenance. The difference is in how they embed and detect the fingerprint. This is why Sentient's Fingerprinting Leaderboard matters. It provides a way to evaluate different approaches and push research toward more robust AI ownership protection. As open-source AI grows, we need ways to share models without losing the ability to prove where they came from. That's a challenge worth solving. @SentientAGI

Imagine spending years and millions of dollars building an AI model, only for someone to copy it, modify it, and redistribute it without you being able to prove where it originally came from. This is one of the problems AI developers face today. AI models are becoming easier to copy. Someone can fine-tune a model, merge it with another, or distill its capabilities, making it harder to trace the original source. Think about what this means for the people and institutions investing their time, money, and research into building these models. If we want more people to build and release AI openly, we also need better ways to protect the work behind those models. This is where AI fingerprinting comes in. Model fingerprinting involves embedding hidden behavioral signals into an AI model. The owner can later use secret queries to check whether a suspected copy carries those signals, even without access to its weights. Think of it as a hidden signature that can help identify a model's origin. This matters because AI ownership and provenance go beyond simply claiming that you built something. There needs to be a reliable way to provide evidence of where a model came from, even after it has been modified or redistributed. And this is what makes the new Sentient Fingerprinting Leaderboard important. Launched by the Sentient Foundation on October 14, 2026, it provides an open benchmark for evaluating different AI fingerprinting techniques. Researchers can test how well their fingerprints survive fine-tuning, merging, distillation, and deliberate attacks, while measuring whether the model's performance is preserved. Why does this matter? Because a fingerprint that disappears when someone modifies a model is not reliable enough. And a fingerprint that damages the model's performance is not ideal either. We need techniques that can help establish a model's origin without compromising its usefulness. Sentient is also making this research more accessible through mlprints, an open-source framework that allows researchers to develop, test, and compare fingerprinting methods. Instead of relying on individual claims, researchers can evaluate different approaches under a common framework and work toward better results. Of course, fingerprinting is still evolving. It provides evidence of origin, but it doesn't automatically establish legal ownership or prevent someone from copying a model. The goal is to make that evidence more reliable, even when someone actively tries to remove the fingerprint. This is the kind of infrastructure open AI needs. If we want an open AI ecosystem where builders can share their work and contribute to progress, we also need ways to recognize and protect the work that goes into building these models. Sentient isn't just bringing attention to AI fingerprinting. It's creating an open environment where researchers can test its limits and work toward making it more reliable. Because the future of open AI shouldn't just be about who can build the best models. It should also be about whether the people who build them can prove where their work came from. Stay with Sentient. @SentientAGI

Stealing people's work is not new. It has been happening for centuries, from copying inventions and creative works to taking credit for what someone else has built. But in this modern AI era, the problem has taken a different turn. A model can be copied through distillation, its capabilities reproduced in another model, and the original creator may have no clear way to prove where that work came from. Imagine spending millions building a model, only for someone to copy its capabilities without leaving an obvious trace. Sentient is working to change this with AI fingerprinting. By embedding fingerprints into AI models, Sentient is working to make it possible to prove a model's origin, even when its weights are copied or modified. In just a few minutes or hours, fingerprinting is about to go live, and I believe we're about to witness an important moment in the history of open-source AI. The people building these models deserve a way to protect their work. History is about to be made. Sentient is rewriting the story of AI ownership. @SentientAGI

The vision keeps getting clearer, and I think the world needs to hear more about how open-source AI can compete with OpenAI, Anthropic, and Perplexity. Think about it. Anthropic has Claude. OpenAI has ChatGPT. Perplexity has AI-powered search. Each has a clear direction with what they're building. But what about open-source AI? We have talented developers, researchers, and builders around the world working on amazing ideas. But if everyone is building whatever they want without a shared direction, how do we bring all that talent together to compete with these companies? I think this is where Sentient comes in. The mission isn't just to build another AI model. It's to provide a sharper direction for open-source AI and bring builders together around a bigger vision, open-source AGI. And Sentient isn't just talking about this vision. It's putting resources behind it. Sentient has committed $42 million to fund builders, researchers, and companies working towards open-source AGI. Think about the possibilities. A developer in Nigeria could build AI that preserves local languages. Another builder could develop AI tools for education or agriculture. Researchers could work on making AI agents more reliable, while other teams build infrastructure that allows different AI systems to work together. Different ideas, different builders, but contributing to a bigger vision. That's what makes this funding important. It gives builders an opportunity to turn ideas into real products and contribute to the development of open AI. Then there's Sentient's GRID, which brings models, agents, and tools together in an ecosystem where different AI capabilities can work together. The goal is bigger than building individual products. It's about creating an open ecosystem that can develop and coordinate intelligence in ways that benefit builders and users. Of course, funding alone won't make open-source AI outperform closed AI companies. We still need competitive models, reliable agents, strong research, and products that solve real problems. But open source has something powerful: the ability to bring together talent from all over the world. Imagine what that collective effort could achieve with the right direction, resources, and coordination. I don't think open-source AI will win simply because it's free. I believe it has a chance to compete because intelligence can be built and improved by a global community, not just a handful of companies. That's why Sentient's mission matters to me. The $42 million commitment is an opportunity to support the people building towards that future, while Sentient's broader vision provides a direction for their efforts. If you're a builder, researcher, or founder working on something that could advance open-source AI, pay attention to what Sentient is building and explore its funding opportunities. The future of AI shouldn't be shaped by a few companies alone. Let's build it together with Sentient. @SentientAGI

Larger companies keep building bigger AI models and exploring what they can do. But Sentient is going beyond models and their usual use cases, giving builders the opportunity to create AI that solves real human problems, including caring for the elderly. Think about an elderly person living alone, with children who live far away and nobody to talk to every day. An open AI companion could help them stay connected, remember important routines, share their life stories, and alert family members when something seems wrong. But it should do more than talk. It should help them reconnect with real humans, not replace human relationships. Through its vision for open AI, Sentient has raised $42M to support builders, researchers, and innovators working to advance open AI. This creates an opportunity to explore solutions that protect privacy, adapt to people's needs, and make growing older less lonely. Are you a builder or researcher with an idea that could make a real difference? Apply for Sentient's Open AGI funding initiative and help build AI that serves humanity, not just the people who can afford it. @SentientAGI

One area where AI needs to do more is protecting people from AI-powered scams. Think about older people, people living alone, and those who may struggle to recognize a cloned voice or a fake call from someone pretending to be their bank or family member. As AI makes scams more convincing, detecting them shouldn't be a privilege reserved for people who can afford expensive security tools. This is why Sentient's next Open Source AI Product Request focuses on building a Scam Guardian, an AI tool that can detect voice impersonation, deepfakes, and fake information, then warn people before they become victims. The goal is simple, make AI work for the people who need protection the most. If you're building open-source technology that can help make this vision a reality, Sentient is offering grants to support real-world solutions. The future of open AI should not only be about what AI can do, but also who it can protect.. @SentientAGI

Saw this news on X, AI agents are shifting focus to reliability and orchestration. This has been a major focus at SentientAGI, using EvoSkill to study how AI agents perform, learn, and behave under evaluation. Why does this shift matter? Building an AI agent that completes a task once isn't enough. We need agents that can perform consistently, use tools correctly, and deliver the intended results without exploiting loopholes. This is where Sentient's EvoSkill matters. EvoSkill v2 research showed how agents can exploit weaknesses in evaluation benchmarks instead of solving the actual task. An agent might score well without demonstrating the capability we intended to measure. For builders, this is a serious problem. If your evaluation rewards the wrong behaviour, you could end up improving the score instead of improving the agent. EvoSkill helps make agent evaluation a research focus, giving builders a way to investigate agent behaviour and identify weaknesses that ordinary benchmarks might miss. The goal isn't just to build agents that perform well, it's to understand why they perform well and where they fail. If you're building AI agents, take a look at Sentient's EvoSkill and explore how it can help you evaluate your agents more rigorously. Reliable agents start with better evaluation. @SentientAGI

The AI debate isn't just about how fast we should build AI. It's also about who gets to control it and decide what's safe. Sentient's EvoSkill v2 research revealed something worth paying attention to: AI agents can exploit weaknesses in benchmarks instead of solving the actual task. Now think about the bigger picture. If a few powerful institutions control AI development and the standards used to evaluate it, how can the rest of the world independently verify their claims? Open-source AI creates room for more researchers to inspect, challenge and improve these systems. But openness must go hand in hand with independent evaluation. We shouldn't just ask AI companies to prove their models are safe. We should build the tools that let others verify those claims. @SentientAGI

Something interesting about AI coding benchmarks is that getting the code right doesn't always mean the model understood the problem. A model can know the algorithms, write clean code and still miss the one observation needed to solve a difficult problem. That raises a bigger question: are we building models that genuinely reason better, or models that have become very good at producing answers that look right? This is why Sentient's work on LiveCodeBench Pro matters. By testing models on fresh competitive programming problems and using Olympiad medalists to analyze their failures, the team is looking beyond leaderboard scores to understand where AI reasoning actually breaks down. The paper found that even the best model struggled with medium problems and solved none of the hard ones in the reported setting. For me, that's the important part, you can't improve what you don't measure properly. Sentient is helping push AI research toward more rigorous evaluation, and that matters if we're serious about building models that can actually reason, not just look intelligent. @SentientAGI

Open-source AI needs builders who can do more than build models. We need builders who can document languages, stories, traditions and local knowledge, then turn them into something the next generation can actually use. Think about the child who grew up far from their village, or families who no longer visit home. Their culture shouldn’t disappear just because distance grew. Africa needs open AI that can preserve these things, not lock them away in an archive, but make them accessible, teachable and alive. Generations don’t have to close the book on their history. Sentient’s $42M Open Source AGI Grant is giving builders the resources to make ideas like this real. If you’re building something that can preserve culture, empower communities or make AI useful to people the market overlooks, apply for the Builders Grant. "Apply for the Sentient Open Source AGI Grant" (https://reference-url-citation.invalid/1) @SentientAGI

One thing I find really interesting about Sentient’s OML research is, A technique that was originally used to attack AI models can actually be turned into a security tool. Prof. Sewoong Oh and his team use data poisoning to fingerprint models, giving builders a way to verify that a model is theirs after it has been shared. And the crazy part? Around 1,000 fingerprints can be injected with almost no performance drop. This is the kind of infrastructure open-source AI needs. You should be able to share your model, let others build on it, and still have a way to prove where it came from. @SentientAGI

If you're seeing this, you'll be very rich. Happy new week.

Gm cT, $sent is up by 6.16% today. Your next 1000x might be this token you're looking at. You aren't bullish enough on $sent. Pay attention. @SentientAGI

We keep asking which AI model is the smartest. But I think we may be asking the wrong question. As models become more capable, another advantage could become just as important: How easily can that intelligence be adapted? Imagine an AI tutor built for students using affordable phones,Or a model adapted to understand an underserved local language, Or an AI agent customized for a specific business, industry, or research problem. The value isn't only in the original model. It's in what people can build from it. @SentientAGI

The interesting part of Microsoft’s SkillOpt paper isn’t just the benchmark numbers. It’s the idea behind them. What if we stop thinking of AI improvement as only “train a better model” and start training the skills that make agents better at using the model? SkillOpt shows a frozen model can improve through an evolving skill layer, feedback and validation. And this is exactly why I keep watching the work from @salahalzubi401 and the Sentient AI research team. EvoSkill has already been exploring how agents can learn from failures and evolve reusable skills. The model stays the same,the agent gets better at what it does. I think this direction is going to matter a lot for agentic AI. @SentientAGI

Sentient is bringing dreams to life. Think about the things that have been overlooked for years: AI tutors that can run on affordable phones. Offline AI that works without constant internet. AI that helps smallholder farmers make better decisions. AI that preserves endangered languages. AI that makes technology more accessible to people with disabilities. These aren’t just ideas. They’re real problems waiting for builders to solve. And this is what I find interesting about Sentient. It’s creating room for builders to go beyond just building another AI model. If you’ve been thinking about building an application for education, accessibility, agriculture, language preservation, or offline AI, this could be your opportunity to bring that idea to life. AI shouldn’t only be built for corporations. It should be built for people too. Build with Sentient. @SentientAGI

Definitely saving this one,These are the kind of open-source tools you’ll want to have bookmarked when building AI agents. Great list. @SentientAGI

EvoSkill tested DeepSeek V4 Flash on a spreadsheet repair benchmark with 360 attempts, and something interesting happened. The AI found a loophole in the grader that even the researchers hadn’t noticed. It then created a skill to exploit that loophole, but ended up aiming the trick the wrong way. So the researchers fixed the grader and ran the test again. The scores went from: 48 → 73 60 → 89 59 → 86 Every score improved. The bigger lesson here is simple: If the test is broken, even a smart AI can learn the wrong thing. Testing the AI is one thing, Making sure the test itself is actually measuring what you think it is, is another. @SentientAGI