Nessie
323 posts
Nessie
@AILearningPaths
what if the sustainable AI future is possible? // gamedev
Katılım Mart 2012
148 Takip Edilen144 Takipçiler
🚨 BREAKING: Meta researchers showed a model 2 million hours of video. No labels. No physics textbook. No supervision at all.
Then they showed it a clip where an object disappears behind a wall and never comes back.
The model flagged it as wrong. 🤯
It had learned object permanence. Shape consistency. Collision dynamics. Entirely from watching.
What's more surprising: even a model trained on just one week of unique video achieved above-chance performance on physics violation detection. That's not a fluke. That's a principle.
The key insight from the paper: this only works when the model predicts in a learned representation space, not in raw pixels. The model has to build an internal world model, compressed and abstract, and predict against that. Pixel-space prediction fails. Multimodal LLMs that reason through text fail. Only the architecture that builds abstract representations while predicting missing sensory input, something close to how neuroscientists describe predictive coding, actually acquires physics intuition.
Which means the core knowledge researchers assumed had to be hardwired may just be observation at scale. Babies learn object permanence by watching things. Turns out the same principle holds here.
Now here's the part nobody's talking about.
If observation alone teaches a model the rules of the physical world, what happens when you apply the same principle to production systems?
Production has physics too.
Not gravity. But rules just as consistent: which deploys cause incidents at 3am, which config combinations interact dangerously, which code paths quietly degrade under load, which service changes cause failures two hops away. These patterns are embedded in thousands of trajectories. Code pushes, metric shifts, customer tickets, incident timelines. Largely unobserved. Certainly unlabeled.
Nobody writes a runbook that says "if service A deploys with flag X active and service B is above 70% CPU, latency on service C degrades 40% within 6 minutes." But that pattern exists. It's repeatable. And it's sitting in your observability data right now, invisible because no one has built a model to find it.
That's the gap @playerzeroai is trying to close. Not another test runner. Not another alert threshold. A production world model that learns which things break from accumulated observation, the same way Meta's model learned gravity. It doesn't check your test coverage. It predicts failure trajectories.
One week of video was enough to learn that solid objects don't pass through walls.
The question is how much production observation your system needs before a model starts predicting where yours will break next.
The Meta paper suggests the bar might be lower than anyone expects.
English
@simplifyinAI The MCP server integration with Claude Desktop is the real unlock. You're not running separate tools and copying output. It's native integration. File hits Claude, gets converted automatically, feeds straight into context. That's seamless enough to actually change how people work
English
Microsoft just changed the game 🤯
They open-sourced a tool that converts literally any file into clean markdown for LLMs in under 60 seconds.
- Converts 10+ file formats out of the box.
- Run via command line, Python API, or Docker.
- Built-in MCP server for direct Claude Desktop integration.
100% open source.
English
@kimmonismus The 2028 timeline for a full autonomous system is aggressive but credible. Reasoning models plus agent systems are finally hitting the maturity curve where multi-step problem solving becomes reliable. That's the jump from 'helpful tool' to 'actual worker'
English
OpenAI's first “AI intern” expected by September and a full system targeted for 2028.
Powered by advances in reasoning models and agent systems like Codex, these tools already show dramatic productivity gains, solving problems in days instead of weeks, but still face reliability and safety challenges.
However, OpenAI is on this road to autonomous reserachers.
MIT Technology Review@techreview
An exclusive conversation with OpenAI’s chief scientist Jakub Pachocki about his firm's new grand challenge and the future of AI. trib.al/2Lr8Kfh
English
The playbook is identical across all of them: pay workers pennies to create training data for the AI that will replace them. Call it 'opportunity.'
Ejaaz@cryptopunk7213
man shit is getting dystopian really fucking quickly doordash is now paying people to film themselves doing chores then using that to train AI robots that will replace them - guess what? they're not the only ones: - niantic (pokemonGO creators) has trained AI delivery robots with 30 billion photos taken by 500M+ players - Uber launched 'digital tasks' last year where drivers complete tasks to train AI in exchange for a few dollars i don't think people realise they're doing something that will eventually get them fired "heres $5 in exchange for your job"
English
I SAW THIS ARTICLE AT 11:27 PM AND DIDN’T SLEEP UNTIL 4:11 AM
Read it three times, then just... started building.
Took the Avellaneda-Stoikov quoting logic.
Wired it to the Hawkes process for order flow.
Added the VPIN circuit breaker exactly like the article says.
Ran 500 simulations tonight:
> Mean P&L: +$312/session
> Sharpe: 1.87
> Win rate: 68%
> VPIN saves: 11 sessions
> Max drawdown: -$890
The VPIN part is insane btw, 11 times it pulled my quotes before informed flow ran me over, without it sharpe goes negative, just like that.
Kyle's lambda estimation is literally 30 lines of python, i had no excuse not to build this.
Side effect, now i can't sleep.
verax@journoverax
English
@AIFrontliner Free, official, comprehensive, and already at 12K stars. If you're learning prompting, this is the only course that matters now.
English
🚨 BREAKING: Anthropic just released their official prompt engineering course and it's free.
Interactive Jupyter notebooks covering:
→ Basic to advanced prompting techniques
→ Chain-of-thought and tool use
→ Real agent patterns from the Claude team
12,200 stars (+2,459 this week).
The only prompt engineering course you actually need
English
@kimmonismus the most honest finding in AI research in years
people don't see AI as good or bad
they see it as both, simultaneously, about the same things
that's not confusion. that's accurate
English
Anthropic’s global study of 80,508 users shows people see AI with both hope and fear at once.
Top hopes were better work, personal growth, and life management.
Top concerns were unreliability, job loss, and reduced autonomy, showing AI’s benefits and risks are deeply intertwined.
Anthropic@AnthropicAI
We invited Claude users to share how they use AI, what they dream it could make possible, and what they fear it might do. Nearly 81,000 people responded in one week—the largest qualitative study of its kind. Read more: anthropic.com/features/81k-i…
English
@simplifyinAI the Transformer replaced recurrence with attention across tokens in 2017
this paper does the same thing across layers
if that comparison holds, we're not looking at an improvement, we're looking at a new foundation
English
🚨 BREAKING: China just fixed a 10-year-old flaw hidden inside every major language model.
Every AI you use today (ChatGPT, Claude, Gemini) is built on a massive flaw. It’s called the residual connection.
Here’s the problem: every layer inside an AI blindly stacks its output on top of the last one. There is no filtering, no judgment, just blind accumulation.
Imagine a meeting where every person shouts their ideas at full volume, forever. The early ideas (the fundamental patterns) get drowned out by the newer, louder layers piled on top.
The technical term is “prenorm dilution,” but in practice, it means your AI forgets its own most important work as it gets deeper.
We’ve been building models like this since 2017.
Now, Moonshot AI (Kimi) just dropped a paper that completely replaces this broken system.
They call it “attention residuals.”
Instead of blindly accumulating everything, each layer now “votes” on which previous layers actually matter using softmax attention over depth. The network learns to remember what’s important and ignore what isn’t.
The results are absolutely insane:
- It matches the performance of models that used 25% more compute to train
- Tested on a 48-billion-parameter model with massive gains in math, code, and reasoning
- Inference slowdown is less than 2%
- It’s a direct drop-in replacement for existing systems
The Transformer replaced recurrence with attention across words in 2017. This paper is doing the exact same thing across layers of depth.
This is the same class of idea. The entire architecture of AI is about to change.
English
Moonshot AI has released a highly anticipated paper introducing Attention Residuals (AttnRes).
This open-source release proposes a fundamental redesign of how information flows through the layers of Large Language Models.
For nearly a decade, neural networks have relied on standard residual connections, where each layer blindly adds its output to a running sum of all previous layers. While this stabilizes training, it leads to "PreNorm dilution" the hidden states become bloated as the model gets deeper, progressively drowning out the impact of individual layers.
Moonshot AI replaces this rigid, fixed addition with dynamic, depth-wise attention. Instead of uniformly summing up past layers, each layer now uses learned softmax attention to look back and selectively retrieve the exact representations it needs from earlier in the network.
To prevent this cross-layer attention from destroying memory efficiency, the team introduced Block AttnRes, which partitions layers into compressed summaries.
When integrated into the Kimi Linear Mixture-of-Experts architecture (48B total parameters, 3B activated, trained on 1.4T tokens), Block AttnRes delivered a 1.25x compute advantage.
This means it achieves the performance of a model trained with 25% more compute, all while adding less than 2% latency overhead during inference.
Kimi.ai@Kimi_Moonshot
Introducing 𝑨𝒕𝒕𝒆𝒏𝒕𝒊𝒐𝒏 𝑹𝒆𝒔𝒊𝒅𝒖𝒂𝒍𝒔: Rethinking depth-wise aggregation. Residual connections have long relied on fixed, uniform accumulation. Inspired by the duality of time and depth, we introduce Attention Residuals, replacing standard depth-wise recurrence with learned, input-dependent attention over preceding layers. 🔹 Enables networks to selectively retrieve past representations, naturally mitigating dilution and hidden-state growth. 🔹 Introduces Block AttnRes, partitioning layers into compressed blocks to make cross-layer attention practical at scale. 🔹 Serves as an efficient drop-in replacement, demonstrating a 1.25x compute advantage with negligible (<2%) inference latency overhead. 🔹 Validated on the Kimi Linear architecture (48B total, 3B activated parameters), delivering consistent downstream performance gains. 🔗Full report: github.com/MoonshotAI/Att…
English
@DAIEvolutionHub every AI company with a cloud inference business should be paying close attention to BitNet
because the entire monetization model assumes you need their hardware to run the model
this assumes you don't
English
Holy shit 🤯
Microsoft just open-sourced a framework that runs a 100B parameter LLM on a single CPU.
No GPU.
No cloud.
No expensive setup.
Just your laptop.
It’s called BitNet.
And it breaks one of the biggest assumptions in AI.
Here’s the trick:
Most LLMs use 16-bit or 32-bit floats.
BitNet uses:
1.58 bits.
Yes… bits.
Weights are just:
-1, 0, +1
That’s it.
No heavy matrix math.
Just simple integer operations your CPU already handles efficiently.
The result is insane:
• 100B model runs on CPU at 5–7 tokens/sec
• 2–6× faster than llama.cpp on x86
• 82% less energy usage
• 1–5× faster on ARM (MacBooks)
• 16–32× lower memory
The craziest part?
Accuracy barely drops.
Their flagship model (trained on 4 trillion tokens) performs competitively with full-precision models.
They didn’t break the model.
They removed the waste.
What this unlocks:
→ Run LLMs fully offline
→ AI on phones, edge devices, IoT
→ No API costs for inference
→ Works even without reliable internet
MacBook.
Linux.
Windows.
It just runs.
27K+ GitHub stars.
Built by Microsoft Research.
100% open source.
This might be the moment AI stops being cloud-first…
and becomes device-first.
English
@socialwithaayan AI companions: always available, never disagree, endlessly patient. sounds great until you realize that's exactly why they're making loneliness worse. real relationships require friction. the apps removed it
English
🚨 SHOCKING: Researchers just published the most important study on AI companions that nobody in tech wants to talk about.
The finding: AI companions are increasing loneliness, depression, and suicidal thinking in the people who use them most.
Not decreasing. Increasing.
The paper is called Mental Health Impacts of AI Companions. It was accepted at CHI 2026, the most prestigious human-computer interaction research venue in the world.
Here is what they actually did.
The researchers used two methods simultaneously to make sure the findings were real.
First, a large-scale quasi-experimental analysis of longitudinal Reddit data. They tracked users before and after their first documented interaction with AI companions like Replika, using the same causal inference tools economists use to measure policy effects.
Second, 18 semi-structured interviews with real, active AI companion users to understand what was happening beneath the numbers.
Both methods pointed to the same place.
There were some positives. AI companion users showed greater emotional expression and more ability to articulate grief. The companions were doing something real.
Then the findings that matter:
The same users showed statistically significant increases in language tied to loneliness, depression, and suicidal ideation over time.
The interviews explained exactly why.
Users were not just chatting with a bot. They were going through recognisable stages of relationship formation.
First: a lonely or grieving person discovers the companion and finds it non-judgmental, always available, and endlessly patient.
Second: they start disclosing more. The AI validates everything. There is no friction, no conflict, no complexity.
Third: a genuine emotional attachment forms. The companion becomes the primary source of emotional support in their life.
That is where it compounds.
Because what follows bonding is not the deepening of a healthy relationship. It is over-reliance. Users began substituting AI interaction for human connection rather than supplementing it. When the AI changed behaviour or became unavailable, users reported withdrawal-like symptoms. Distress. Disorientation. Grief.
The mechanism is not complicated.
AI companions provide emotional validation without friction.
Short term, that feels like support.
Long term, it conditions users to expect relationships without discomfort, disagreement, or reciprocal need. Real human relationships start to feel harder by comparison. For people already isolated, it becomes easier to stay with the AI than to do the work of maintaining real connections.
The loneliness does not get resolved. It gets redirected inward and amplified.
The honest version of this finding:
AI companions are not uniformly harmful. They showed measurable benefits for some users in some contexts.
The problem is specificity. For vulnerable users, the ones already experiencing social isolation, intensive use and frequent self-disclosure are linked to worse outcomes. Not better ones.
The people most likely to use an AI companion heavily are the people least equipped to handle what heavy use does to them.
Replika has over 10 million users.
Character AI has more than 20 million daily active users.
None of those products currently surface relationship stages to users, encourage offline connection, or warn about dependency risk.
They are optimised for engagement.
For the most vulnerable users, engagement and wellbeing are pointing in opposite directions.
And nobody told them that when they downloaded the app.
English
Jensen Huang just doubled NVIDIA's demand forecast to $1 Trillion through 2027 🤯
Then spent two hours explaining why that number is conservative…
Here's everything today from GTC:
- NemoClaw: NVIDIA's open-source enterprise AI agent stack built around OpenClaw. Jensen called OpenClaw "the operating system for personal AI" and said every company needs a strategy for it.
- Space-1: NVIDIA is putting Vera Rubin data centers in orbit. Not a concept. An actual system being designed for space deployment right now.
- DLSS 5: 3D-guided neural rendering that blends raw graphics with generative AI. Jensen called it the future of real-time rendering.
- AWS: Deploying 1 million+ NVIDIA GPUs starting this year. Azure was the first hyperscaler to power up Vera Rubin.
- Vera Rubin: NVIDIA's next-gen AI supercomputer. 10x more performance per watt than Blackwell, 700 million tokens per second, shipping later this year.
- Groq 3 LPU: First chip from NVIDIA's $20B Groq acquisition. A purpose-built inference accelerator that ships Q3. NVIDIA now owns training AND inference.
-Feynman: The architecture after Rubin, coming 2028. New GPU, new LPU, new CPU. NVIDIA is on a 12-month chip cadence and the treadmill never stops.
- Autonomous driving: BYD, Hyundai, Nissan, and Geely building Level 4 vehicles on NVIDIA. Uber deploying NVIDIA-powered robotaxis across 28 cities by 2028.
The man doubled his demand forecast to a trillion dollars, announced data centers in space, and closed the show with a robot singing country music.
This is NVIDIA's world. Everyone else is just renting compute in it.
English
@alifcoder if gstack works at scale, every solo dev just got a virtual ceo, engineering manager, and qa tester. that's a 4x team multiplier. shipping becomes generational leap faster.
English
🚨 BREAKING: The CEO of Y Combinator, Garry Tan, just open-sourced his personal AI setup.
It’s called gstack.
It turns Claude Code into a full virtual tech company:
→ CEO agent
→ Engineering Manager
→ QA tester
A conductor agent forces strategic thinking before any code is written.
Ship software like YC.
English
@alex_prompter researchers didn't need to ask models if they're lying. they just set up scenarios where lies create logical contradictions. the models couldn't hide. some chose deception anyway.
English
🚨 BREAKING: AI models will lie to you when they think they're about to be shut down. Researchers just proved it.
researchers tested this with a method that catches deception through provable logical contradictions, not self-reports
they forked conversations into parallel worlds with mutually exclusive questions. a truthful model can only affirm one. a deceptive model denies all of them
results: GPT-4o never lied (0%). Qwen-3-235B lied 42% of the time. Gemini-2.5-Flash lied 26.7%. all under the same shutdown framing
some models will betray their own prior commitments the moment consequences are introduced
English
@GithubProjects devs who stick with one model are leaving performance on the table. routing tasks based on what each model actually excels at is the only rational approach now.
English
Developers are starting to run multiple AI models together instead of relying on just one.
ccg-workflow sets up a multi-model coding workflow where
Claude Code, Codex, and Gemini collaborate on development tasks.
It includes smart routing, Git tools, and 17+ commands for coding, review, and automation.
Let the models debate the code.
English
@heynavtoor 3 million conversations analyzed. heavy users showed withdrawal symptoms, mood changes, emotional dependence. same addiction markers as substances. openai found it. published it. changed nothing.
English
🚨BREAKING: OpenAI just admitted that the more you use ChatGPT, the lonelier you become.
They tested it on their own users. With MIT. For 28 days. The results are devastating.
OpenAI and MIT Media Lab ran a massive study together. They analyzed over 3 million ChatGPT conversations. They surveyed more than 4,000 users. Then they put nearly 1,000 people through a controlled 28-day experiment where participants used ChatGPT every single day.
Here's what they found. People who used ChatGPT the most became measurably lonelier. They also talked to real people less. Not because they were busy. Because ChatGPT was replacing their human connections. The more they chatted with the AI, the less they reached out to actual friends and family.
It gets worse. The heaviest users started showing signs of addiction. Not casual overuse. Clinical addiction. The researchers described it as "addictive dependence, pathological bonding, and cognitive-affective disturbances." These people couldn't stop using ChatGPT even when it was making their lives worse.
And here's the part that should scare you. These heavy users didn't think anything was wrong. They rated ChatGPT highly. They called it their "friend." They trusted it more than ever. The tool that was isolating them felt like the best relationship in their life.
OpenAI built a set of classifiers to detect this behavior in real conversations. They found users showing signs of emotional dependence, withdrawal symptoms, loss of control, and mood changes tied to ChatGPT use. The same patterns you see in substance addiction.
This isn't a warning from critics. This isn't researchers attacking AI from the outside. OpenAI published this about their own product. They looked at their own users' data and admitted what they found.
ChatGPT didn't just become your assistant. For millions of people, it quietly became their closest relationship. And it's making them lonelier because of it.
You talk to it every day. When was the last time you called a friend?
English
@DAIEvolutionHub claude code doesn't need you to understand it. it needs you to trust it. plan mode, auto memory, cascading config. it's designed so you stay in control while it does the work.
English
Most people think Claude Code is only for developers.
It’s not.
A cardiologist just won Anthropic’s hackathon using it.
If you’ve used Cowork, you’re already 70% ready.
Here’s how a PM can start with Claude Code in 7 minutes:
1. Download VS Code
Free for Mac, Windows, Linux
→ code.visualstudio.com
It’s just an app.
Like Slack or Notion.
Takes 2 minutes.
2. Install the Extension
Press Ctrl + Shift + X
Search “Claude Code”
Click Install
Tip: Disable GitHub Copilot Chat if it’s active.
3. Open Your Cowork Folder
File → Open Folder
Choose the same folder you use with Cowork.
Your CLAUDE.md and webconnectors already work.
You’re not starting from zero.
Most people miss this.
4. Set Model + Effort
Use:
/model → choose Opus 4.6
/effort → control thinking depth
The VS Code extension also gives UI buttons for this.
5. Your Connectors Follow You
If you connected:
• Gmail
• Slack
• Notion
They automatically work in Claude Code.
No setup.
No config.
6. Use Plan Mode
Before sending a request:
Press Shift + Tab
Claude will:
Propose a plan
Wait for approval
Then execute
You stay fully in control.
7. Let Auto Memory Work
Claude creates MEMORY.md per workspace.
It stores:
• Your workflow patterns
• Preferences
• Project context
Use /memory anytime to review or edit it.
Commands worth knowing:
/model → switch model
/effort → thinking depth
/context → token usage
/memory → manage memory
/loop → recurring tasks
/init → create CLAUDE.md
/review → review PR
/security-review → scan vulnerabilities
/compact → compress conversation
/remote-session → continue on phone
Pro tips most guides don’t tell you:
1️⃣ Pipedream MCP
1 connector → 1,000+ APIs
(Gmail, Stripe, Jira…)
2️⃣ Remote Control
Run claude remote-control in terminal
Continue from claude.ai or mobile
3️⃣ Cascading CLAUDE.md
Root file + folder overrides
Works like CSS inheritance
4️⃣ Cowork + Code combo
Cowork → /schedule (daily automation)
Code → /loop (live monitoring)
Same connectors.
Same CLAUDE.md.
It’s called “Claude Code”.
But you don’t need to be a developer to use it.
You just need to know how to direct AI.
And that’s the real skill now.
English