The_Year_of_Six_Revelations
1
2
3
4
5
6
7
8
9
10
11
Slide 1
NotebookLM to PPTXTurn NotebookLM Slides into Editable PPTXExport PPTX
Log In
A Retrospective on the Paradigm Shifts That Redefined Al
2025: The Year of Six Revelations
NotebookLM
Based on the analysis of Andrej Karpathy
NotebookLM
RLHF
SFT
PRETRAINING
Aligning the model with
human preferences through reinforcement learning.
Human Feedback (RLHF):
3.
follow instructions by imitating human examples.
Supervised Finetuning (SFT): Teaching the model to
2.
Building the vast knowledge foundation.
Pretraining:
1.
For years, the path to a powerful LLM followed a stable, three-stage formula established since 2020:
At the Dawn of 2025, the Recipe Was Proven
compute from pretraining to RL runs. The inflection point was OpenAl's o3 model, where the difference became intuitively obvious.
This new, longer optimization stage shifted
Supporting Detail**
**
RLHF
SFT
PRETRAINING
RLHF
PRETRAINING
NotebookLM
A fourth stage emerged: Reinforcement Learning from Verifiable Rewards (RLVR). Instead of imitating human examples, models were trained against objective rewards in environments like math or code.
SFT
RLVR
The result was profound: models spontaneously developed strategies that look like “reasoning” - breaking down problems and finding novel solutions.
Revelation I: The Engine Was Rebuilt From the Inside Out
NotebookLM
"Human neural nets are optimized for survival.. LLM neural nets are optimized for imitating humanity's text, collecting rewards in math puzzles, and getting that upvote."
LLM INTELLIGENCE (Optimized for Verifiable Tasks)
HUMAN INTELLIGENCE (Optimized for Survival)
We began to internalize the "shape' of LLM intelligence. It is not a well-rounded, biological intelligence optimized for survival. It's a 'jagged' intelligence, optimized for narrow, verifiable tasks. It is at once a genius polymath and a confused grade schooler
Revelation Il: We Aren't Raising Animals.We're Summoning Ghosts.
BASE MODEL
NotebookLM
Offer the user an "autonomy slider."
4.
Provide a human-in-the-loop GUI.
3.
2.
Engineer context for the model.
1.
*Analogy: LLM labs produce the "generally capable college student." LLM apps organize and train them into "deployed professionals."*
Orchestrate complex chains of LLM calls.
The rise of apps like Cursor proved the existence of a thick, valuable “LLM app" layer. These are not just wrappers; they are orchestrators that perform four key functions for specific verticals:
Revelation Ill: A New Application Layer Solidified
NotebookLM
longer a website you visit, but a 'little spirit/ghost that 'lives' on your computer.' This is a distinct, new paradigm of interaction.
This was a critical insight, as the intermediate state of AI requires close human collaboration, not remote cloud orchestration.
The AI is no
The Core Idea:
Claude Code demonstrated a new paradigm: the AI agent that lives on your computer. By running locally, it gains access to your private environment, data, and context.
Revelation IV: The Al Moved Onto Your Machine
*Example: Karpathy vibe-coded a custom BPE tokenizer in Rust without needing to deeply learn the language.*
on technology diffusion, empowering regular people first. For professionals, it makes countless small, bespoke, or singleuse programs worth creating. Code becomes "free, ephemeral, malleable, discardable after single use."
It flips the script
Impact:
NotebookLM
In 2025, AI crossed a threshold enabling anyone to build impressive programs simply by describing them in English. This is “Vibe Coding."
Revelation V: The Language of Creation Became Conversation
End of repert.
1ice Lino o report code (84So2seoe--fcrootcteaf7)); End of repert.
1tr0.t0et15808/6G107cbe508e15125158S97166345583$ec557319119F713].
1tDo Dorpart Isss.
dcte Dotoea. t5e.
deta 4EL0 0o[bt5 18$17$880 dOLa Dotnts S0os5GSss
...1CRt4LLrG t6IT OT4rE.GEEE.ItC169es0TOsxC015010LeG6T01S5etoTIT0sCaaD0e1GT0RSScSfbacQ1AGTOEcTe coeT66t8c5ze0ea4tce1cocc0cq50c00c1000cd10c05a00rc0a400E540c56cct1610G0cerrc8t0cc50c0cCatc4cEcoco5 J06c5408cLDe00cc0.0cGzCc0Gq8GoccCr050GG6cto6a405005Co50La0G500Gc056cG0cc5c45GGcGac0GcC00G5c58GG5e 104C00506005196651000E48196L1160E66860864644081C61E606G406400060864G064560D6G6G0C004500446564046 308045662006r04460660004068000400006400604508908606760/665000060016406045600466006000066056603646 008060₺6410800000446050661406066050866660520046666846066066100806864T6086606466016000006680646676 66660₹E68404536006560662680006668c866666616666006646685566616600446000683000004666080606046616626 #660006c400460a00606400006400d560c0555008800060c0504666e40000640040a50000506086440566c80006604606 563006E59011656001604605569066668C0560€5546520056553001642006206061042086500156000045626046004666 15GC00604000$540004$1000554665061000555604060064155404066001654500666000555401600285$040046160006 30et0050ec6465060&E00080500c0c662ce5550c08564c0c558c058atcet050501tc1ce0558t6580ees4oGc8t6c60cc0 45650c150261053161060E19576060101587091660701350758580404t7960050$c660Y55000005c010701840750. bo8tc15c00e4s.500104r00500:St8_r0066t
Rocprt ceeerorce.Oate soctvocc!
--foreatvtext
GENERATE_REPORT
NotebookLM
Key Text: Humans consume information visually and spatially. LLMs must learn to speak our language: images, infographics, and animations. Google's Gemini Nano Banana was the first hint of this future, jointly using text, image generation, and world knowledge to "show," not just "tell".
Analogy:“Chatting" with LLMs is like issuing commands to a computer console in the 1980s. Text is the computer's favored format, but not the human's.
THE FUTURE:VISUAL INTERFACE
THE PAST:COMMAND LINE
el
Revelation VI: The GUl for LLMs Has Arrived
NotebookLM
2025 left us with a paradox. LLMs are a new kind of intelligence—jagged, alien, and powerful. They are simultaneously capable of breathtaking brilliance in verifiable domains and baffling incompetence in others. This is the nature of the ghost we have summoned.
The New Reality: Smarter Than We Imagined, Dumber Than We Expected
NotebookLM
- Andrej Karpathy
rapid and continued progress and that yet there is a lot of work to be done."
*I simultaneously (and on the surface paradoxically) believe that we will both see
Key Text: "I don't think the industry has realized anywhere near 10% of their potential even at present capability."
The Field is Wide Open
NotebookLM
A cinematic retrospective based on '2025 LLM Year in Review' by Andrej Karpathy.
Strap in.