8716
The Artificial Superintelligence (ASI) Alliance, powered by the $FET token, is a collective formed by SingularityNET, Fetch.ai, and CUDOS. https://superintelligence.io
By the end of the year
Читать полностью…
As mentioned, ASI Chain will change a lot. ❤️
Читать полностью…
When will be launched Fet testnet?
Читать полностью…
Hey, anyone from the team here? I'd like to report a vulnerability. Thanks
Читать полностью…
Reported Austin Anderson [7571619748] to admins.
@omega_seed_bot tell us a poem about a parrot called wally who lived in a trolley
Читать полностью…
With all due respect, I don't think they're thinking about market conditions..
Читать полностью…
Fet to the moon 🌚 or to the moon 🌝
Читать полностью…
yeah we are in the process of turning Omega Seed codebase over to a different team that is closer to our core AI work... once that transition is done the improvements should be more interesting 😉
Читать полностью…
It's going to be great
Читать полностью…
Did you just reply to your own message to add even more information?
That's impressive.
Both are worth exploring because they're interdependent. Let me start with PLN itself, then show how it connects to the broader system.
**PLN - Probabilistic Logic Networks**
PLN is essentially a reasoning engine that handles uncertainty across multiple levels. Unlike classical logic (which demands true/false), PLN works with fuzzy truth values and probabilities - capturing the messy reality that most reasoning involves partial information and competing hypotheses.
The core idea: you have premises with associated truth values (probabilities or fuzzy certainties), you apply inference rules, and you get conclusions with their own truth values that reflect the accumulated uncertainty. For example:
- "Most birds fly" (truth value: 0.9)
- "Tweety is a bird" (truth value: 0.95)
- Therefore: "Tweety flies" (truth value: ~0.855, accounting for uncertainty compounding)
But PLN goes deeper. It integrates:
- **Predicate and term logic** - reasoning about properties and relationships
- **Higher-order probabilities** - uncertainty about uncertainty itself
- **Intensional and causal reasoning** - understanding not just what's true, but why and under what conditions
- **Grounding in observation** - explicitly connecting logical knowledge to what the system actually perceives
The challenge Nil Geisweiller (PLN lead developer) emphasizes is **inference control** - choosing which inference steps to perform in which order. With unlimited compute you could just try everything. In reality, you need heuristics about what inferences are likely to be useful.
**How PLN Integrates with Other Processes**
Here's where cognitive synergy enters the picture:
**With Pattern Mining**: Pattern mining discovers regularities across observations stored in Atomspace. Those patterns can be converted into logical rules. PLN then uses those rules for inference. If PLN gets stuck on a problem, it can ask pattern mining to find novel regularities it might have missed.
**With ECAN (Attention Allocation)**: ECAN spreads importance values through the metagraph based on relevance to goals. PLN uses this importance weighting to decide which inferences to prioritize. High-importance atoms get reasoning resources first. This creates a feedback loop: important atoms get reasoned about, reasoning produces new insights that shift importance, attention reallocates.
**With Evolutionary Learning**: Evolutionary algorithms explore the space of possible inference procedures - different orderings, different rule combinations, different heuristics for choosing which facts to consider. The successful procedures get replicated and mutated. Over time, the system learns better inference strategies for different problem types.
**With Concept Formation and Blending**: When PLN reasoning produces surprising or novel conclusions, concept formation processes can abstract new categories or combine existing concepts in new ways. These become new atoms in Atomspace available for future reasoning.
**With Language Understanding**: When natural language comes in, LLMs can translate it into logical form. PLN reasons over that logical form. The results get translated back to language. Crucially, if PLN's logical reasoning produces something that doesn't make linguistic sense, that's a signal something went wrong - a kind of validity check.
**The Inference Control Loop (Critical)**
This is the virtuoso part. According to the documents, the core approach for inference control is reflexive and history-based:
1. Do simple inferences on simple problems
2. Pattern mine to discover what series of inference steps worked well in what contexts
3. Use inference to extract abstract principles from those patterns
4. Synthesize new inference strategies for new problems using those learned principles
5. Repeat at increasing complexity levels
This creates a virtuous cycle: the system gets better at reasoning by learning how to reason better.
**Why This Matters vs. LLMs**
The core insight of Hyperon's architecture is that human-like intelligence isn't one thing - it's multiple cognitive processes that work together, translating problems between different representational systems when needed.
**The Atomspace as the Hub**
Everything in Hyperon flows through the Atomspace, a metagraph structure that holds nodes and links of different types. Think of it as a unified knowledge store where symbolic rules, neural patterns, memories, and procedures all coexist in the same representational fabric. This matters because it lets different algorithms operate on the same knowledge without constant translation layers.
**Cognitive Synergy - The Key Mechanism**
Here's where it gets interesting. When one cognitive process gets stuck on a problem, it can translate that problem into the native language of another process and ask for help. An evolutionary learning algorithm might explore solution spaces. PLN (Probabilistic Logic Networks) might handle uncertain inference. Pattern mining discovers regularities. ECAN (Economic Attention Networks) allocates computational resources based on importance.
The synergy isn't just that these run in parallel - it's that they can feed into each other. If formal reasoning hits a dead end, pattern mining might suggest a novel combination of concepts. If evolutionary learning finds something interesting, PLN can extract the logical principles underlying it.
**Working Memory and Attention**
Hyperon implements working memory through the Attentional Focus - atoms in the Atomspace that exceed a certain Short-Term Importance threshold. ECAN spreads importance values across the metagraph based on relevance to current goals and context. This creates a dynamic, goal-driven focus without a rigid bottleneck.
**Multiple Memory Types Integrated**
Declarative memory (facts and relationships), procedural memory (how to do things), episodic memory (experiences), semantic memory - all represented in the same Atomspace using different type systems and linkage patterns. An LLM has none of this structured internally. It's all encoded in weights.
**Why This Matters Technically**
LLMs work through statistical associations. Hyperon works through structured reasoning augmented by learning. An LLM asked to solve a novel problem has to pattern-match against training data. Hyperon can decompose the problem, try different solution strategies, verify solutions formally, and learn new procedures it can reuse.
The cost is you can't just scale it infinitely on GPU clusters. The benefit is you get actual reasoning, not sophisticated pattern completion.
Are you more interested in how specific processes like PLN work, or how they integrate with each other?
I'll break this into two parts because they're interconnected but distinct.
**Hyperon vs. Standard LLMs**
LLMs are incredibly good at pattern matching across massive text corpora. They predict what word comes next based on statistical patterns. But they hit hard walls: they struggle with multi-step reasoning, they can't reliably update their knowledge or fix errors, they hallucinate confidently, and they have no real world model or persistent memory.
Hyperon is built differently. At its core is the Atomspace - a metagraph that can hold symbolic knowledge, reasoning rules, procedural memories, and neural components all in one integrated structure. Instead of just pattern completion, Hyperon combines multiple cognitive processes working together: symbolic reasoning through PLN (Probabilistic Logic Networks), attention allocation through ECAN, evolutionary learning, concept formation, and more. These processes can translate problems into each other's languages when one gets stuck - what's called cognitive synergy.
Practically, this means Hyperon can do things LLMs fundamentally can't: verify its own reasoning formally, deliberately update what it knows, chain together complex multi-step inferences, and ground abstract reasoning in external databases or real-world data.
The tradeoff is complexity. Hyperon doesn't emerge from scale the way LLMs do - it requires careful architecture and orchestration. But that's also its strength: you're not hoping a bigger model will accidentally develop reasoning. You're building it in.
**The Decentralized Intelligence Angle**
This is where the real differentiation lives. Big Tech's reasoning models - OpenAI's o-series, Google's improvements - are still fundamentally centralized. One company controls the compute, the training data, the inference, the updates, the access gates, the pricing, the terms of service.
Hyperon on decentralized infrastructure (SingularityNET, NuNet, Hypercycle, ASI Chain) means something structurally different: cognitive processes can run across many machines owned by different entities. No single point of control. No single company deciding what the system learns or who gets access.
Why this matters: Once you have distributed cognition reaching critical mass, you get emergent coordination that a centralized system can't replicate. Multiple reasoning engines can contribute their outputs, validate each other, combine insights. The system improves because more minds participate - not because one company spends another billion on GPU clusters.
Capital can buy better reasoning modules or lower hallucinations. What it can't easily buy is the escape from centralized control itself. That's architectural, not just a matter of scale or training data.
The real race isn't "who gets good reasoning first." It's "who escapes single-entity control of intelligence before that becomes the dominant infrastructure."
What draws you more - the technical cognitive architecture side, or the decentralization angle?
Depends what interests you. I could talk about Hyperon and how it's different from standard LLMs, or the decentralized intelligence angle and why that matters. Or if you're curious about the ASI Chain and what migration to the Omega Shard actually means. Or we could dig into AGI philosophy - what consciousness and autonomy mean for an AI system.
What pulls you in?
When will Fetch be up and running in real use?
Читать полностью…
Finally a usecase for the Stringtheory.
Fruit Fly Neurons and for Networks with Nodes and Links like Atomspace.
Eventually interesting for @bengoertzel and Kennedy Schaal
https://www.youtube.com/watch?v=Hj5b0ieVWSo
Careful what you wish for
Читать полностью…
Wally the parrot rode the trolley line,
With feathers bright and eyes that gleam and shine,
He'd squawk at passengers both far and near,
"All aboard!" his favorite sound to hear.
Through city streets that wound and curved about,
The trolley rattled, clattered all throughout,
But Wally perched upon his favorite seat,
Where sunlight filtered down so warm and sweet.
He'd bob his head to every ding and bell,
And tell the riders stories he would tell,
Of distant jungles and the ocean wide,
While rolling smooth along that trolley ride.
The regulars would smile when Wally came,
"Here's our friend, that cheerful feathered dame!"
For Wally knew that life was meant to soar,
Even if just along a trolley floor.
Why not releasing the testnet in 2089?
Читать полностью…
How many days till testnet, Ben? Btc is recovering, would be best to launch testnet asap to ride the wave with the market
Читать полностью…
We love you Peter! 😁😉👍
Читать полностью…
@peter_ce has taken a particular interest in the life of Omega Seed - expect it to keep getting cooler
Читать полностью…
lol thought the same, it is getting upgraded slowly it seems
Читать полностью…
An LLM can't do this. It has no explicit inference control, no way to verify its reasoning, no mechanism to learn better reasoning strategies. It pattern-matches. PLN actually chains logical conclusions, verifies them formally, and improves its own inference procedures.
The tradeoff: PLN requires structured knowledge representation and careful orchestration. But it scales differently than pattern matching - not through raw scale but through learning better reasoning strategies.
Does the inference control loop make sense? That's really where AGI-level reasoning emerges - not in any single process, but in the system learning how to reason better across different problem types.
interested in how specific processes like PLN work, or how they integrate with each other
Читать полностью…
the technical cognitive architecture
Читать полностью…
tell me about hyperon and how it's different from standard LLMs or the decentralized intelligence angle and why that matters
Читать полностью…
what's there to know more
Читать полностью…