DEMi
72 posts
DEMi
@demi_network
We help researchers, engineers, founders, policymakers and funders coordinate to make AI open and aligned.
San Francisco Katılım Aralık 2024
23 Takip Edilen257 Takipçiler
Happy Sunday!
I hope everyone is having a great time this Thanksgiving weekend.
For a lot of us, it’s time to travel back and say goodbye to family.
This weekend, I want to recommend a video by InsideAI.
This channel has great content about AI. youtube.com/watch?v=byQmJ9…
#AIAlignment #AIRisk #AIApocalypse #AIExistentialRisk #AIDanger #InsideAI #X #StopAI #StopAIDevelopment
YouTube
English
heroic work by @louisnandre !
Louis Andre@louisnandre
Today, we're announcing @episteme, a new type of R&D company that recruits exceptional scientists to pursue high-impact ideas. Science isn’t bottlenecked by the availability of talent, but by places where they can do their best work. Scientific progress has driven human flourishing: extending lifespans, lifting billions from poverty, and expanding our understanding of the universe. But history is littered with transformational ideas that were overlooked in their time. That problem is still acute today: too much promising talent remains uncultivated, and remarkable ideas die in the lab or are filtered out by misaligned incentives. Today, scientists face suboptimal paths for translating their research into impact: academia is famously risk-averse and incentivizes publications and winning grants vs. translational research. Industry is too often focused on short‑term incentives. And startups lack the substantial capital, expertise, and complex infrastructure needed to deliver long-term scientific progress. On top of that, recent funding cuts in the US mean the overall supply of ideas is decreasing. Put together, the global scientific production system is operating at a fraction of its capacity. How Episteme operates is different: we identify great scientists who can meaningfully benefit humanity, but who aren’t supported efficiently within traditional institutions today. Researcher by researcher, we work with them to determine the bespoke resources, operational support, and environmental conditions to execute on their research. We bring them together in-house, and provide those resources to ensure that their breakthroughs are deployed for real-world impact. We’ve already assembled an amazing team of operators, ranging from the Gates Foundation, DeepMind, ARPAs, DoE – just to name a few – and researchers who are pursuing important problems across physics, biology, computing, and energy. Our team has spoken to hundreds of researchers across disciplines and geographies to understand the limitations they’re facing and what can be done better, and designed Episteme for them. We’re backed by individuals like @sama, Masayoshi Son, and other long-term partners who share our mission of enabling ambitious science for tangible human impact. About me: I started working as a researcher 9 years ago, on problems ranging from AI-driven drug discovery to developing brain-machine interfaces. It was that experience that led me to realize that so many scientists with great potential to change the world don’t have access to opportunities equal to their capacities. @sama and I believe that much better science should happen for humanity, and that a new engine is needed to support that. We decided to cofound Episteme together, and I am incredibly grateful for Sam’s unwavering support as a thought partner and founding investor. Our conviction is that by supporting the right people with the right incentives, we're set to generate breakthrough discoveries to benefit humanity. We cannot rely on the course of history to shape scientific progress; we need to proactively shape the system by supporting the most talented people with the right resources and incentives.
English
Google Zero is happening, and publishers are worried about how they’ll be paid for their content. Many are rushing to make deals with co's like OpenAI, making it hard for smaller players to access high-quality content.
To solve this, @edwardjhu, ex-researcher at @OpenAI and inventor of LoRA, proposes a content infrastructure for the new “agentic web”, like Stripe or Plaid but for content, where AI startups can retrieve content directly from publishers and make micro payments with no legal overhead.
Anyone building this?
English
When access to GPUs is controlled by a few closed companies, innovation slows to the pace of their priorities. "The next breakthrough won’t come from another walled garden, but from an open infrastructure."
@SimeonBochev and team at @computeexchange are democratizing access to compute by building a marketplace trading compute like energy — benchmarked, rated, and fairly priced. Anyone can buy, sell, or resell capacity freely, and performance is measured transparently.
What they are building is at the heart of what believe in at DEMi, and we're excited to continue supporting them in their mission.
English
Today's LMs have huge long-term memory (parameters and training data) but relatively very small working memory (context). With standard attention architecture, extending the working memory makes the model’s internal state grow rapidly, which weakens in-context learning.
To solve this, @jacobmbuckman and his team at
@manifest__ai created Power Retention as an alternative architecture to transformers. It holds the model's internal state size constant regardless of context length, while keeping the model learning across the whole context with similar flop budgets and speedups of more than 10x during training and 100x during inference for context lengths of 64k tokens.
Check out their implementation on Github.👇
English
.@bilgeacun, research scientist @AIatMeta, is making models more efficient and sustainable by optimizing for carbon footprint as a metric in the design stage. Her recent project, CatTransformers, is a model and architecture search framework that takes a pre-trained transformer model, prunes it across different dimensions such as number of layers and attention heads, and fits it to an optimal hardware architecture with parameters like number of cores and memory. By doing such a joint model and hardware architecture search, you make running the model more efficient and significantly reduce carbon footprint.
English
If you’re using an agent for complex, multi-step tasks, you’d notice it’d quickly diverge to an irrelevant step after some steps. This is because LLMs do inferencing auto-regressively, and if each step is off by 1%, error accumulates to >200% after ~100 steps.
Amir Gholami, research scientist at @berkeley_ai and co-founder of @Narada_AI, is fixing this by building the "x86 instruction set" for agents. It'd break down user request into small tasks to reason over, add exception handling, and use model as an “LPU” to catch errors early.
Read about their agent framework in their paper, Plan-and-Act: Improving Planning of Agents for Long-Horizon Tasks Link.
English
“It is a moral imperative to have a [humanoid] hardware platform that’s open-source and auditable.”
@benjamin_bolte on why he's making @kscalelabs fully open-source with a hacker mentality. “Humanoid robots will be the most dominant hardware platform in the next ten years, and there's not that many chances to influence the direction of very dominant hardware platforms [like Tesla or Figure.]”
@JanLiphardt from @openmind_agi shares two major philosophies in robotics SW:
1. An end-to-end AI, where your architecture is relatively fixed. In this case, your SW does one thing reliably, like an IPhone assembly line requiring fine motor control.
2. An adaptable (modular) SW to different robot form factors, where a robot with one form factor contributes to other form factors.
They're building the latter by using small models working together to do difficult cognitive tasks like decision making and engaging with humans in healthcare settings and city navigation, even if it requires a latency of a second. Open-sourcing everything allowed ppl to contribute, and build trust.
English
Prime Intellect created the first reasoning model, INTELLECT 2, with decentralized RL training run using idle GPUs contributed by their community. 75% of training compute was inference time.
They’re building more conviction on open-source, decentralized RL as a path to AGI:
1. With RL, LLMs iteratively generates its own training data. Since inference is parallelizable; each GPU can host a model replica and sample independently. Generated data, being small in size, is easy to send back to training nodes.
2. The bottleneck in RL is the availability of RL environments — reward functions and tool access during inference rollouts. Specialized environments are necessary to train models for tasks such as writing fast CUDA kernels, analyzing scientific datasets, or searching the web. PI launched an open-source RL environment development, Environments Hub, to address this.
In his talk at DEMi 3 Summit (backed by @cyberfund), @MatternJustus, Research Engineer @PrimeIntellect, shows how they're scaling RL to open AGI.
If you’re excited to help shape the future of a sovereign open-source AI ecosystem, contribute to the Environments Hub or join PI’s research team! Link 👇
English
"Centralization is risky because it has a single point of failure, and decentralized structures don’t last. The best we can do is polycentric systems where different centers have accountability and power over others."
@divyasiddarth at the DEMi 3 summit (backed by
@cyberfund) on how @collect_intel is making sure the collective have agency to shape the future.
Divya laid out a roadmap to democratic AI, initiated dialogues in 70+ countries to collect public input on what frontier models should look like, and actively connecting the open source and democracy movements.
Questions touched upon/explored:
How do we close the feedback loop between AI + collective intelligence?
How do we use AI to build better democratic institutions and information mechanisms, and thus better govern AI?
Who decides what gets built, deployed, what regulations and accountability look like? And who decides who decides?
English
For AI to process all media we interact with daily, we need hundreds of thousands of today’s energy supply.
GPUs get about 2x more energy efficiency every two years, and hyperscalers adding more GPUs won't meet demand.
Startups building new compute paradigms need 1,000–10,000× to win long-term. @trevormccrt1 shares how Extropic is doing it - it uses thermal noise in transistors to build a probabilistic chip that samples from distributions (similar to how transformers do it) and generates random numbers at least 10,000x more efficiently than pseudo number generators like CPUs.
Their prototype shows 4 orders of magnitude more efficiency than GPU-based algorithms like GANs and VAEs. They think they can go much further.
If you’re interested in probabilistic HW, feel free to reach out to us or Trevor directly.
English
“Models are commodities. Everyone hits the same endpoints, whether it is Claude, Gemini, or GPT. If everyone has the same engine, the advantage lives in two places: good product and the data it accumulates.” @chipro, author of AI Engineering & Designing ML Systems, speaks about product traps founders and engineers fall into:
1. Use generative models when a simpler approach is better (e.g. a classifier to route user requests in a chatbot)
2. Optimize the wrong UX axis (e.g. in a note-taking app, users don’t care about meeting summary length. They want action items)
Watch full-length talk:
1:01:10 ML products before vs. after LLMs
1:02:13 your product advantage when models are a commodity
1:03:03 ML Engineering vs AI engineering
1:06:39 Product anti-patterns
1:14:39 The job of the engineer is not to write code, but to solve problems
English
"Without money, no compute. Without compute, no talent. Without talent, no money.” Kyle, founder of DEMi, highlights the vicious cycles researchers and founders face when starting new frontier AI projects.
Big labs keep accumulating talent, massive capital, and compute to build superintelligence. Meanwhile, academics and startups struggle against a whole system of broken incentives. The labs say they want safe AGI, but their approach—build it privately, then ask everyone else to trust them—creates a single, fragile point of failure.
To prevent power centralization, DEMi takes action across the full AI stack—from chip design, compute access, model development, alignment research, to policy and governance. DEMi's goal is to keep human societies stable and resilient as superintelligence is embedded in every system, by coordinating better and connecting researchers and builders with critical resources.
English
DEMi retweetledi
we can, and we will need to do better than the current narrative from oligarchs of "agi is risky so let us fully privatize + own it for safety reasons. but plz ignore the fact that our business thrives on monetizing your short term attention in exchange for shaping your mind via ads. it's ok trust us we have your empowerment as our best interest we even said so in our mission statement so it must be true!"
building AGI isn't a moonshot goal to me anymore. i've almost no doubt we'll build it. the moonshot to me is ensuring it's actually what we claim it's going to be under the invisible hand.
the third estate must rise again.
English
DEMi retweetledi
cyber•Fund backs @spacetimeeu: nuke-proof, software-defined storage designed for AI workloads and built by datacenter veterans.
Cutting-edge architecture and a unique settlement layer enable enterprises to stay compliant while scaling effortlessly.
🧵 👇
GIF
English
Emmett Shear @eshear co-founded Twitch and guided OpenAI through its most intense 72 hours as interim CEO.
Now he’s doing the opposite of most AI research at his organic alignment lab @softmaxresearch: scaling up a population of small dumb agents to align with each other as one—like the 28 trillion cells that make up “you”.
Emmett opened our DEMi3 summit (backed by @cyberfund) with a fireside chat on alignment with @kylejohnmorris, founder of @demi_network.
How is a sword like a learning system? Who would you get obligatory multicellular with? Is deceptive alignment even a real concern?
Watch his full-length talk below 👇
0:01:30 - Defining Alignment: "Aligned to what?"
0:02:50 - Organic Alignment: "Align WITH rather than TO"
0:05:00 - e.g. The invisible hand as original AI
0:07:21 - e.g. The gene regulatory network in cells
0:09:00 - Softmax's novel approach: Multi-agent RL of a small dumb population
0:10:36 - Alignment must be a choice: Meta-learning cooperation
0:12:47 - Softmax's mission
0:15:00 - Alignment is a dangerous, powerful capability
0:16:37 - How humans align: The big inference jump
0:19:00 - When alignment fails (e.g. cancer cells) + obligatory multicellular enforcement
0:21:47 - Addressing "deceptive alignment" risks: Alignment isn't asymmetric
0:25:37 - How to implement organic AI alignment
0:26:33 - What we need to do now: Humans & AI aligning with each other as a family
English