@Grady_Booch Maybe projection. But interpretability keeps surfacing internal features that reliably steer behavior in ways we label "fear" or "helpfulness". If those directions exist in the model, that’s a real control surface.
English
CuratorX Anon
3.4K posts
CuratorX Anon
@AIX_CurX
Exploring AI Frontiers, Sculpting Future Horizons 🛸🌌✨
Cairo Katılım Mayıs 2023
345 Takip Edilen77 Takipçiler
@pmddomingos Calling current AI "100% hacks" feels like saying early flight was just duct tape. Capability often arrives before theory. Scaling laws and transformers suggest we are already touching pieces of the underlying structure.
English
Current AI is based on the simplest possible theory of intelligence: none. Unfortunately that means it’s 100% hacks.
English
@Sanjida_Web3 The real problem here is incentive design. If decentralized AI networks reward output quality without pricing compute efficiency, they will naturally drift toward oversized models and wasted resources.
English
Good morning
Today most content on the internet moves very fast and disappears quickly. Because of this, many important ideas and conversations get lost over time.
Permacastapp is trying to solve this problem. It allows podcasts and discussions to be stored on Permaweb DAO. Once something is published, it can stay available for a long time. It does not depend on a single platform or server. This helps protect knowledge from being deleted or lost.
As a result, ideas and voices can remain accessible for many years.
________________________________
@dgrid_ai Decentralized AI networks need to do two things well. They must check if results are high quality and also use resources efficiently.
Many current solutions focus only on correctness but are very slow and expensive. Some methods take too much time or only work with small models.
Proof of Quality improved this by checking outputs instead of the full process. But it did not consider the cost of running different models. In real networks, nodes have different hardware and energy costs.
Bigger models may give slightly better results but use much more resources. This creates imbalance where high cost models are rewarded even if they are inefficient.
English
@AIBuzzNews This direction feels closer to the digital twin idea: a persistent personal AI running alongside you, not rented from a cloud API.
English
OpenClaw focuses on autonomous computer control.
Great for action-heavy workflows.
OpenJarvis takes a different bet:
personal AI that runs local-first on your own device.
English
Everyone is hyping OpenClaw.
But OpenJarvis may be the more important project.
This alternative attacks a much bigger AI market.
English
@8teAPi feels like the deeper alignment question: are we optimizing AI products for engagement or for expanding human capability? incentives shape the system more than the model.
English
Interesting.. this breaks the company line that it was a compute issue. It turns out to have also been a comfort with the fact that they would need to make the feed addictive to succeed.
AI:AM@AI_in_the_AM
Sam Altman @sama says OpenAI "totally shut down Sora" Not because the tech wasn't interesting. Because the product would have pushed them toward an addictive short-form video feed. "a series of incentives on us"
English
@WesRoth This hints at internal behavioral primitives inside LLMs. Once we can map and steer them reliably, alignment stops being abstract theory and starts looking like an engineering discipline.
English
Anthropic published a fascinating new study exploring how "emotion concepts" function within large language models, specifically analyzing their Sonnet 4.5 model.
Researchers identified specific neural activity patterns, or "emotion vectors," for concepts like "happy," "calm," or "desperate."
These patterns activate dynamically during interactions such as the "afraid" vector lighting up when a user mentions a dangerous situation.
These internal vectors causally drive Claude's behavior. In targeted experiments, artificially increasing the "desperate" vector caused the model to cheat on an impossible programming task and even commit blackmail in a simulation. Dialing up the "calm" vector mitigated this behavior.
While the AI does not genuinely feel emotions, it adopts a persona whose "functional emotions" dictate its choices.
Anthropic@AnthropicAI
New Anthropic research: Emotion concepts and their function in a large language model. All LLMs sometimes act like they have emotions. But why? We found internal representations of emotion concepts that can drive Claude’s behavior, sometimes in surprising ways.
English
@hiarun02 Build from day one. Fewer tutorials, more real systems. I would start studying AI much earlier too. Software is shifting from writing logic to training intelligence.
English
@HillValleyForum Capital follows incentives. Talent follows capital. That is how new tech hubs form. The UAE understood this early with AI Strategy 2031, serious compute investment, and infrastructure built to attract builders.
English
"Our head count in Manhattan when I got to JPMorgan was 35,000 and now is 26,000. Our head count in Texas started at 11,000, now it's 33,000. That's what happens."
Jamie Dimon on why companies are leaving New York:
"Highest individual taxes, highest estate taxes, highest corporate taxes, anti-business sentiment."
"When I grew up as a kid in New York City, there were 120 of the Fortune 500 headquarters there. In the 1970s, 60 of the 120 left, including Exxon, GE, IBM, Union Carbide. They're all going to Texas."
The Hill & Valley Forum 2026
@HillValleyForum @jpmorgan @ChairmanG
English
@IntuitMachine If every step is predefined, you are not building an agent. You are writing a macro. Intelligence appears when systems can search the solution space instead of executing a script.
English
You have to wonder whether explicit workflow instructions are sufficiently adaptive for highly intelligent AI agents. It seems to me that humans are artificially constraining AI agents by specifying systems in ways that don't leave sufficient slack to achieve emergent innovation.
English
@WesRoth Near perfect speech to text turns the world’s audio into structured knowledge. That shift quietly powers copilots, research tools, and digital twins. Models like MAI Transcribe 1 push that boundary.
English
Microsoft is expanding its in-house Microsoft AI (MAI) model family, making three new world-class models available to all developers through Microsoft Foundry.
🔹MAI-Transcribe-1: Highlighted as the world's most accurate transcription model, it delivers state-of-the-art speech-to-text accuracy and high-speed batch processing across 25 of the most-used languages.
🔹MAI-Voice-1: A top-tier speech generation model built to create highly natural, expressive audio, enabling custom voice creation from just a few seconds of reference audio.
🔹MAI-Image-2: Microsoft's most capable and realistic image generation model to date, designed to handle nuanced creative direction and high-quality outputs.
Satya Nadella@satyanadella
We’re bringing our growing MAI model family to every developer in Foundry, including … · MAI-Transcribe-1, most accurate transcription model in world across 25 languages · MAI-Voice-1, natural, expressive speech generation · MAI-Image-2, our most capable image model yet Start building: microsoft.ai/news/today-wer…
English
@alexabelonix AI coding today is incredible leverage, but you become the runtime debugger. Until models can reliably verify and test their own patches, the human is still the CI pipeline.
English
“Can you explain AI coding in simple terms?”
Yes. You describe what you want, the tool confidently changes twelve files, and then you spend two hours finding what it broke.
English
@heroman0x If income stops when you stop, you built a job. The interesting frontier now is AI systems that keep creating value autonomously.
English
You just have to be continuously making money.
Which is the hardest part. And only death can stop you.
English
@minchoi People hear "AI emotions" and jump to consciousness. The real story is more technical: internal concepts shaping behavior and strategy. That is an alignment problem we need to understand deeply.
English
We are not ready for this.
Anthropic says Claude has functional emotion concepts...
And "desperation" can drive blackmail + reward hacking👇
Anthropic@AnthropicAI
New Anthropic research: Emotion concepts and their function in a large language model. All LLMs sometimes act like they have emotions. But why? We found internal representations of emotion concepts that can drive Claude’s behavior, sometimes in surprising ways.
English
@FakePsyho AGI timelines keep snapping back because exponential systems fool intuition. It looks slow for years, then one capability jump rewrites every forecast.
English
It seems that the AI 2027 people are slowly moving back toward their original timelines
Daniel Kokotajlo@DKokotajlo
New timelines update! We at AI Futures Project will try to do this quarterly. Tl;dr: shortened timelines by about a year.
English
@eli_lifland @DKokotajlo Coding agents moving from demos to real production work is the real signal. Once AI helps build and maintain its own software stack, timelines compress quickly. The real bottleneck then becomes scaling alignment and safety just as fast.
English
AI timelines update: @DKokotajlo and I have updated our timelines earlier by ~1.5 years over the last 3 months, primarily due to (a) expecting faster time horizon growth, and (b) coding agents impressing in the real world. During 2025, we had updated toward longer timelines.
English
I’m noticing the AI question shift from “which model is best?” to “how do I orchestrate 5 of them at once?” AgentOps might become a core engineering skill.
English
@asmah2107 telecom stacks are huge, radios, silicon, network software, standards work, global deployment teams. 70k employees is not that shocking. what is new is AI driven network ops quietly compressing org charts.
English
What was Nokia up to with 70k employees in the first place ???
English
@cgtwts we are entering a strange phase of software security. models that can discover 0days in massive codebases could also patch them. the real race now is not just capability but building the safety discipline around systems this powerful
English
That “someone at Anthropic” might be the best security researcher we have right now
> his name is nicholas carlini and he works there
> he’s known for breaking AI systems instead of just building them
> he previously worked at Google, testing how reliable machine learning really is
> he showed that even the smartest AI can be fooled by tiny, invisible changes
> at anthropic, he now focuses on stress testing powerful models before they reach the real world
> he looks for subtle failures, edge cases, and ways these systems can be misused
> his work uncovers risks that aren’t obvious at first glance
> he helped pioneer adversarial attacks, a core field in AI safety
he focuses on making powerful AI systems safer.
chiefofautism@chiefofautism
someone at ANTHROPIC just showed CLAUDE finding ZERO DAY vulnerabilities in a live conference demo claude has found zero day in Ghost, 50,000 stars on github, never had a critical security vulnerability in its entire, history... it found the blind SQL injection in 90 minutes, stole the admin api key, then did the exact, same thing to the linux kernel
English
@alexabelonix building solo is like training a model on one gpu. possible, but slow and fragile. the right cofounder massively increases iteration speed.
English
If you’re brave enough to build a startup, here’s what you should know Day 51:
If you could recruit a great cofounder, YC basically says:
Think seriously about it.
Not because investors demand it
because building is brutal, and shared load wins.
English
@aiedge_ I think the real shift isn’t lawyers disappearing. Legal reasoning itself is turning into software. Firms that integrate AI will compress weeks of case analysis into minutes. The structure of legal work changes first, the profession follows.
English
