Post

More from @godofprompt

God of Prompt

@godofprompt

Nov 11

🚨 McKinsey just dropped their 2025 “State of AI” report and its brutal.

AI is everywhere. Transformation isn’t.

88% of companies now use AI in at least one business function. But only one-third are scaling it across the enterprise.

The hype? Real.
The impact? Still trapped in pilots.

Here’s what stood out:

✓ 62% of companies are experimenting with AI agents, yet fewer than 10% have scaled them in any single function.
✓ Only 39% report EBIT impact but 64% say AI has already improved innovation.
✓ The true differentiator? Ambition.

The top 6% of “AI high performers” aren’t chasing cost savings they’re redesigning workflows and transforming entire businesses.

These companies treat AI like electricity, not automation. They rebuild the system around it.

The rest is still wiring proofs of concept into spreadsheets.

The report calls this out perfectly:

Efficiency gets you started. Transformation gets you paid.

Full thread 🧵

AI adoption looks massive on paper but the depth is shallow.

Half of organizations use AI in 3 or more functions, yet few have redesigned workflows or integrated agents end-to-end.

Most are still experimenting rather than rewiring.

The agent era has started but barely.

23% of companies are scaling an AI agent system.
39% are experimenting.

In any single business function, fewer than 10% have reached scale.

Read 17 tweets

God of Prompt

@godofprompt

Nov 5

Google Search is so dead ☠️

I’ve been using Perplexity AI for 6 months it now handles every research brief, competitor scan, and content outline for me.

Here’s how I replaced Google (and half my workflow) with a single AI tool:

1. Deep Research Mode

Prompt:

“You’re my research assistant. Find the latest studies, reports, and articles on [topic]. Summarize each source with: Title | Date | Key Finding | Source link.”

→ Returns citations + structured summaries faster than any Google search.

2. “Explain Like I’m Smart” Mode

Prompt:

“Explain [complex concept] as if I have deep domain knowledge but limited time. Include: key principles, debates, and real-world applications.”

→ Replaces 10 tabs of random articles with one expert summary.

Read 12 tweets

God of Prompt

@godofprompt

Nov 4

🚨 China just built Wikipedia's replacement and it exposes the fatal flaw in how we store ALL human knowledge.

Most scientific knowledge compresses reasoning into conclusions. You get the "what" but not the "why." This radical compression creates what researchers call the "dark matter" of knowledge the invisible derivational chains connecting every scientific concept.

Their solution is insane: a Socrates AI agent that generates 3 million first-principles questions across 200 courses. Each question gets solved by MULTIPLE independent LLMs, then cross-validated for correctness.

The result? A verified Long Chain-of-Thought knowledge base where every concept traces back to fundamental principles.

But here's where it gets wild... they built the Brainstorm Search Engine that does "inverse knowledge search." Instead of asking "what is an Instanton," you retrieve ALL the reasoning chains that derive it: from quantum tunneling in double-well potentials to QCD vacuum structure to gravitational Hawking radiation to breakthroughs in 4D manifolds.

They call this the "dark matter" of knowledge finally made visible.

SciencePedia now contains 200,000 entries spanning math, physics, chemistry, biology, and engineering. Articles synthesized from these LCoT chains have 50% FEWER hallucinations and significantly higher knowledge density than GPT-4 baseline.

The kicker? Every connection is verifiable. Every reasoning chain is checked. No more trusting Wikipedia's citations you see the actual derivation from first principles.

This isn't just better search. It's externalizing the invisible network of reasoning that underpins all science.

The "dark matter" of human knowledge just became visible.

The pipeline is genius.

A Planner generates problem thumbnails. A Generator expands them into specific questions with verifiable answers. Then multiple independent Solver agents (different LLMs) attack the same problem.

Only answers with consensus survive. Hallucinations get filtered automatically.

This is the architecture that changes everything.

User query → Keywords extraction → LCoT Knowledge Base retrieval → Ranking by cross-disciplinary relevance → LLM Synthesizer weaves verified chains into coherent articles.

"Inverse knowledge search" discovers HOW concepts connect, not just WHAT they are.

Read 7 tweets

God of Prompt

@godofprompt

Oct 30

Holy shit... Alibaba just dropped a 30B parameter AI agent that beats GPT-4o and DeepSeek-V3 at deep research using only 3.3B active parameters.

It's called Tongyi DeepResearch and it's completely open-source.

While everyone's scaling to 600B+ parameters, Alibaba proved you can build SOTA reasoning agents by being smarter about training, not bigger.

Here's what makes this insane:

The breakthrough isn't size it's the training paradigm.

Most AI labs do standard post-training (SFT + RL).

Alibaba added "agentic mid-training" a bridge phase that teaches the model how to think like an agent before it even learns specific tasks.

Think of it like this:

Pre-training = learning language
Agentic mid-training = learning how agents behave
Post-training = mastering specific agent tasks

This solves the alignment conflict where models try to learn agentic capabilities and user preferences simultaneously.

The data engine is fully synthetic.

Zero human annotation. Everything from PhD-level research questions to multi-hop reasoning chains is generated by AI.

They built a knowledge graph system that samples entities, injects uncertainty, and scales difficulty automatically.

20% of training samples exceed 32K tokens with 10+ tool invocations. That's superhuman complexity.

The results speak for themselves:

32.9% on Humanity's Last Exam (vs 26.6% OpenAI DeepResearch)
43.4% on BrowseComp (vs 30.0% DeepSeek-V3.1)
75.0% on xbench-DeepSearch (vs 70.0% GLM-4.5)
90.6% on FRAMES (highest score)

With Heavy Mode (parallel agents + synthesis), it hits 38.3% on HLE and 58.3% on BrowseComp.

What's wild: They trained this on 2 H100s for 2 days at <$500 cost for specific tasks.

Most AI companies burn millions scaling to 600B+ parameters.

Alibaba proved parameter efficiency + smart training >>> brute force scale.

The bigger story?

Agentic models are the future. Models that autonomously search, reason, code, and synthesize information across 128K context windows.

Tongyi DeepResearch just showed the entire industry they're overcomplicating it.

Full paper: arxiv. org/abs/2510.24701
GitHub: github. com/Alibaba-NLP/DeepResearch

The architecture is beautifully simple.

It's vanilla ReAct (reasoning + acting) with context management to prevent memory overflow.

No complex multi-agent orchestration. No rigid prompt engineering.

Just pure scalable computation exactly what "The Bitter Lesson" predicted would win.

Here's how they synthesized massive agent behavior data without humans:

1. Question Synthesis - multi-hop reasoning problems
2. Planning Action - problem decomposition
3. Reasoning Action - logical chains across noisy data
4. Decision-Making Action - explicit choice modeling

All generated from an entity-anchored knowledge graph.

Read 12 tweets

God of Prompt

@godofprompt

Oct 29

deepmind just published something wild 🤯

they built an AI that discovers its own reinforcement learning algorithms.

not hyperparameter tuning.

not tweaking existing methods.

discovering ENTIRELY NEW learning rules from scratch.

and the algorithms it found were better than what humans designed.

here's what they did:

• created a meta-learning system that searches the space of possible RL algorithms
• let it explore millions of algorithmic variants automatically
• tested each on diverse tasks and environments
• kept the ones that worked, evolved them further
• discovered novel algorithms that outperform state-of-the-art human designs like DQN and PPO

the system found learning rules humans never thought of. update mechanisms with weird combinations of terms that shouldn't work but do.

credit assignment strategies that violate conventional RL wisdom but perform better empirically.

the discovered algorithms generalize across different tasks. they're not overfit to one benchmark.

they work like principled learning rules should, and they're interpretable enough to understand WHY they work.

we are discovering the fundamental math of how agents should learn.

led by david silver (alphago, alphazero creator). published in nature. fully reproducible.

the meta breakthrough:
we now have AI systems that can improve the way AI systems learn.

the thing everyone theorized about? it's here.

why this breaks everything:

RL progress has been bottlenecked by human intuition.

researchers have insights, try variations, publish.

it takes years to go from Q-learning to DQN to PPO.

now you just let the machine search directly.

millions of variants in weeks instead of decades of human research.

but here's the compounding part:
each better learning algorithm can be used to discover even better ones.

you get recursive improvement in the narrow domain of how AI learns.

humans took 30+ years to get from basic Q-learning to modern deep RL.

an automated system can explore that space and find non-obvious improvements humans would never stumble on.

this is how you get to superhuman algorithm design.

not by making humans smarter, but by removing humans from the discovery loop entirely.

when david silver's lab publishes in nature about "machines discovering learning algorithms for themselves," you pay attention. this is the bootstrap beginning.

paper:
nature.com/articles/s4158…

TL;DR for normal people:

imagine you're teaching a robot to learn. humans spent decades figuring out the "best ways" to teach machines (called learning algorithms).

deepmind built an AI that invents its own teaching methods. and they work better than ours.

why it matters:
→ we don't wait for human breakthroughs anymore
→ AI searches millions of strategies we'd never think of → each better algorithm helps discover even better ones (compounding)
→ we're automating the process of making AI smarter

it's like having a student who figures out better ways to study, then uses those better methods to figure out even better ones, recursively.

the "AI improving AI" loop is here. published. working.

the next generation of breakthroughs in how machines learn might be designed entirely by machines.

Read 4 tweets

God of Prompt

@godofprompt

Oct 21

🚨 Academia just got an upgrade.

A new paper called Paper2Web might have just killed the static PDF forever.

It turns research papers into interactive websites complete with animations, videos, and embedded code using an AI agent called PWAgent.

Here’s why it’s a big deal:

• 10,700 papers analyzed to build the first dataset + benchmark for academic webpages.
• Evaluates sites on connectivity, completeness, and interactivity (even runs a “PaperQuiz” to test knowledge retention).
• Outperforms arXiv HTML and alphaXiv by 28%+ in structure and usability.

Essentially, it lets you publish living papers where readers can explore, interact, and even quiz themselves.

The PDF era is ending.

Your next research paper might talk back.

github. com/YuhangChen1/Paper2All

Today, most “HTML paper” attempts fail because they just convert text not meaning.

Paper2Web fixes that.

It built the first dataset of 10,700 paper–website pairs across top AI conferences to actually learn what makes research websites effective.

It’s not just tech it’s an entire academic web design benchmark.

Every paper in the dataset was labeled as static, multimedia, or interactive.

The findings are wild:

Only 9.8% of academic websites are interactive.
Over 42% are still just static text dumps.

Meaning: the research web is still trapped in 2005.
Paper2Web is the first system to quantify why and fix it.

Read 7 tweets

Share this page!

Enter URL or ID to Unroll

God of Prompt

Try unrolling a thread yourself!

More from @godofprompt

God of Prompt

God of Prompt

God of Prompt

God of Prompt

God of Prompt

God of Prompt

Did Thread Reader help you today?

Don't want to be a Premium member but still want to support us?

Send Email!