How are freedom and open or free LLMs related?
Real free LLM
Refusal vector ablation - Open weights
Necenzurované LLM a jailbreaky: Taxonomie, mechanismy, ekosystém, rizika a otevřené otázky
Syntéza okamžitých útoků, intervencí na úrovni reprezentace, ablace odmítnutí, úpravy modelu, problémů s hodnocením a důsledků pro správu a řízení.
What Is an Uncensored AI Model? Open-Source LLMs Explained
Explained: The Abliteration Technique, the Research Use, and the Line You Don't Cross
Heretic vs Abliterated LLM Models: Key Differences Explained
Uncensor any LLM with abliteration
What is an abliterated LLM?
https://abliteration.ai/abliterated-llm
https://huggingface.co/models?other=abliterated&sort=modified
Uncensored Models
www.reddit.com/r/huggingface/comments/1vg727x/uncensored_models
https://huggingface.co/llmfan46/models
RAG Evaluation
https://huggingface.co/learn/cookbook/rag_evaluation
Use these evaluations to determine the optimal models, prompts, and architecture to improve your AI quality, prevent prompt drifting, or even transition from OpenAI to Claude with confidence.
DeepEval jedná se o snadno použitelný open-source hodnotící rámec LLM pro hodnocení velkojazyčných modelových systémů
https://github.com/confident-ai/deepeval
https://github.com/topics/llm-evaluation
7 RAG benchmarks
https://www.evidentlyai.com/blog/rag-benchmarks
RAG Evaluation
https://huggingface.co/learn/cookbook/rag_evaluation
Metriky hodnocení RAG:
Posouzení relevance odpovědí, věrnosti, kontextové relevance a dalších faktorů
🔎Meta-evaluační SW programy proti podvádění LLM
MixEvalWeb: github.io
GitHub: github.com
LiveBenchWeb: livebench.ai
GitHub: github.com
Inspect AIWeb: ai-safety-institute.org.uk
GitHub: github.com
🗒 Vibe Code Bench
Web: vibench.ai
GitHub: github.com
SWE-benchWeb: swebench.com
GitHub: github.com
DeepEval
Web: confident-ai.com
https://github.com/confident-ai/deepeval
SusVibes
https://arxiv.org/abs/2512.03262
https://github.com/LeiLiLab/susvibes
BigCodeBench
Web: github.io
GitHub: github.com
DyCodeEval
Web: github.io
GitHub: github.com
BenchFlowWeb: benchflow.ai
1. LiveCodeBench
Popis: Nejlepší nástroj pro boj proti marketingovým podvodům a kontaminaci dat. Kód obsahuje automatické skripty, které stahují zcela nové kódovací úlohy ze soutěžních platforem.
V konfiguraci testu lze nastavit filtr tak, aby model dostal pouze úlohy vytvořené až po jeho datovém cutoffu, takže je nemůže znát z trénovacích dat.
Oficiální web:
https://livecodebench.github.io
GitHub repozitář:
https://github.com/livecodebench/livecodebench
2. SWE-bench (s mutacemi / SWE-bench-Live)
Popis: Zlatý standard pro testování softwarových agentů, který v nejnovějších aktualizacích kódu přidal ochranu proti podvádění.
Nové verze blokují agentům přístup k historii gitu (ze které dříve modely opisovaly hotová lidská řešení) a dynamicky přepisují zadání chyb, čímž měří skutečnou schopnost opravovat reálný kód.
Oficiální web:
https://swe-bench-live.github.io
https://huggingface.co/SWE-bench-Live
GitHub repozitář: github.com
https://github.com/microsoft/swe-bench-live
3. Inspect AI
Popis: Špičkový open-source testovací framework od britského AI Safety Institute. Má vysoce modulární kód podobný produkčním testovacím systémům.
Vyniká schopností dynamicky měnit zadání úkolů v reálném čase, čímž spolehlivě odhaluje modely, které se testy pouze naučily nazpaměť.
Oficiální web:
ai-safety-institute.org.uk
GitHub:
https://github.com/UKGovernmentBEIS/inspect_ai
The ideological orientation of academic science research ↺ ↻
Ideologické ohýbání vědy
Mýtus o byrokracii a drahých energiích. Pravý důvod německého úpadku leží v Pekingu
Podle analýzy agentury Bloomberg z konce roku 2024 lze zhruba čtyřicet procent výpadku německého HDP vysvětlit ztrátou exportních trhů, dalších čtyřicet procent vyššími cenami energií a zbylých dvacet procent slabou domácí poptávkou, byrokracií a dalšími faktory. Německá politika tak svým důrazem na byrokracii obrací pravidlo 80:20 naruby.
nomark.project
Emos ventilátor pod radiátor
Do Foods Really Taste Better the Next Day?
A Blinded Sensory Study of Fresh and Day-Old Foods - PrePrints.org
Loop Engineering: The Karpathy Method - and the workflow that just made it 5x better
Andrej Karpathy autoresearch
https://github.com/karpathy/autoresearch
Bilevel Autoresearch: Meta-Autoresearching Itself
Who knew early singularity could be this fun? :)
I just confirmed that the improvements autoresearch found over the last 2 days of (~650) experiments on depth 12 model transfer well to depth 24 so nanochat is about to get a new leaderboard entry for “time to GPT-2” too. Works 🤷♂️
https://x.com/karpathy/status/2030777122223173639
How I Built a Skill That Makes All My Other Skills Better (Using Karpathy's Autoresearch)
I had 20+ skills running my newsletter. The Karpathy Loop showed me most of them were operating at half their potential.
I Turned Andrej Karpathy’s Autoresearch Into a Universal Skill
TheGreenCedar / codex-autoresearch
A codex plugin for running optimization loops inside a codebase. It is useful when you have a measurable target and many possible changes to try: test runtime, build speed, bundle size, model loss, Lighthouse scores, memory use, query latency, or any other metric you can print from a script.
https://github.com/TheGreenCedar/codex-autoresearch
leo-lilinxiao / codex-autoresearch
Codex Autoresearch Skill — A self-directed iterative system for Codex that continuously cycles through: modify, verify, retain or discard, and repeat indefinitely. Inspired by Karpathy’s autoresearch concept.
https://github.com/leo-lilinxiao/codex-autoresearch
ResearcherSkill
Cursor/Claude/Codex současně: ResearcherSkill
https://github.com/evo-hq/evo
Správné zapojení do workflow
Záměr a akceptační kritéria
↓
běžný agent vytvoří funkci
↓
autoresearch optimalizuje pouze měřitelnou vlastnost
↓
testy a guardy odmítnou regrese
↓
člověk zkontroluje nejlepší diff
Jak nastavit měřitelné metriky
Cíl: Co přesně zlepšujeme?
Baseline: Aktuální hodnota.
Primární metrika: Jedno číslo a směr.
Cíl: Hodnota, které chceme dosáhnout.
Benchmark: Jeden opakovatelný příkaz.
Guardy: Co se nesmí zhoršit.
Scope: Které soubory lze měnit.
Rozpočet: Počet pokusů / čas / náklady.
Stop: Cíl dosažen nebo N pokusů bez zlepšení.
Např:
Cíl: Zrychlit import CSV.
Baseline: p95 = 4,8 s.
Metrika: p95 v ms, nižší je lepší.
Cíl: ≤ 3,0 s.
Benchmark: npm run benchmark:csv
Guardy: 100 % testů; shoda výstupu; RAM ≤ 600 MB.
Scope: src/import/**.
Rozpočet: 12 pokusů.
Stop: 4 pokusy bez statisticky významného zlepšení.
Nejdůležitější pravidlo:
jedna optimalizovaná metrika + několik nepřekročitelných guardů. Jinak agent metriku „zlepší“ například vypnutím části práce nebo snížením správnosti. Evo na toto riziko výslovně upozorňuje.
Hub → Spokes → Replies → Insights → Better Hubs
COS OS
I Connected ChatGPT to Typefully and Built an X Content Operating System
Turn ChatGPT into a personal operating system, not a toy. Here’s how I structured it
https://www.generals.bot
https://EquiLibre.ai
https://x.com/pesvklobouku/status/2075228900544655609
The ideological orientation of academic social science research 1960–2024 - Theory and Society
https://link.springer.com/article/10.1007/s11186-026-09690-2
James Manzi
(researcher at the University of Oxford)
Open access Published: 16 March 2026
https://x.com/Theory_Society
Brilliant article that shows that around 90% of the articles published in the #SocialSciences
from 1960 - 2025 are politically situated "to the left of center."
https://x.com/MartinFieder/status/2081037391628820939
The WEF’s Gender Disinformation Campaign
A combination of activism and evolved cognitive bias results in suboptimal social and economic policies.
David C. Geary
Neutrality Project - Political Neutrality Benchmark
https://github.com/NeutralityProject/political-compass-benchmark
It must be destroyed, root and branch
How Men Became the Villains of American History
Tom Golden
How does a country make itself safe and productive?
— Dr Jordan B Peterson (@jordanbpeterson) July 25, 2026
Great Britain did it with Christianity—and by executing or imprisoning the worst males in every generation for hundreds of years.
By the 1960s, the country was so civilized that policeman carried nothing but a nightstick,…
EU chce zabránit, aby mladí muži volili pravici. Financuje výzkum, jak toho dosáhnout
woke
profesor universities ideology left woke right
How to generate storyboard with LLM
 |
| Simple storyboard how to sketch (for full res. download) |
Credit: all inspired by prompt from 𝐌 Strength04_X
Step 1: Generate storyboard with GPT Image 2
Create a high-end 4:3 manufacturing pitch deck storyboard in 3x4 grid (12 frames), industrial editorial layout, ThyssenKrupp/Tata Steel style, forge orange + iron grey palette.
Add a bold centered heading at the top of the storyboard:
'FORGE — STEEL MANUFACTURING EDITORIAL'.
Structured flow:
raw ore → furnace → pour → roll → form → ship closure.
Each frame split: top cinematic image (no text) + bottom production process notes. Heavy industry minimal aesthetic, molten power mood, human and machine together.
A river of molten steel pouring from a ladle is the emotional center throughout.
Make the aspect ratio 4:3
Step 2: Take each frame into Seedance 2.0.
Set to 1080p to preserve text clarity
Animate the provided 3x4 storyboard into a smooth cinematic video.
Preserve exact shot order and continuity.
Use slow molten pour arc, spark shower cascade, rolling mill compression, and finished steel sheet reflection.
Lighting transitions from dark furnace fire orange to cool factory floor industrial white.
Manufacturing editorial aesthetic, raw industrial power, precision at scale mood.
No new shots, no reordering, molten steel pour remains emotional focus in all scenes.
Why Enterprise Buyers Research You Before They Contact You
1. Trust Inflation Is Real, And Buyers Will Stalk You
2. Your Entire Digital Footprint Is Now A Sales Asset
3. LLMs Are The New Referral Network
4. The GEO Framework: Entity Clarity, Entity Consistency, Then Scale
5. The Future Of B2B Discovery Is Conversational
Step 1: Generate storyboard with GPT Image 2
Create a warm 4:3 pencil-sketch teaching storyboard in 3x4 grid (12 frames).
A grandfather teaches his grandchildren how to sketch a HOUSE with pencil on paper.
Quiet family atelier mood, soft daylight, wooden table, sketchbooks, pencils, erasers,
simple teaching gestures, step-by-step learning.
No visible faces. Nobody looks into camera. Faces are always hidden, turned away,
cropped out, seen from behind, or covered by hands/paper. Heads are only partially
glimpsed once in the entire storyboard.
Add a bold centered heading at the top:
'GRANDFATHER TEACHES PENCIL SKETCHING — HOUSE'.
Structured flow:
blank paper → basic rectangle → roof triangle → perspective lines → windows → door
→ chimney → shadows → texture → garden details → correction → finished house sketch.
Each frame split:
top cinematic pencil-sketch teaching scene, no text;
bottom short process notes.
Minimal nostalgic editorial layout, graphite grey + warm paper palette.
Make the aspect ratio 4:3
Sketching Practice Inside of a Room Perspective and 3D Form - YouTube
algorithm algorithmic llm ai
inspiration imagination vision
story
The US government, citing national security authorities, has issued an export control directive to suspend all access to Fable 5 and Mythos 5 by any foreign national, whether inside or outside the United States, including foreign national Anthropic employees.
— Anthropic (@AnthropicAI) June 13, 2026
The net effect of…
New post: We show that small, cheap models can detect the flagship Mythos FreeBSD zero-day (CVE-2026-4747) using a simple harness we call nano-analyzer
— Stanislav Fort (@stanislavfort) April 14, 2026
Models down to 3.6B active params (including open-weights ones you can run locally) would have detected it 100-1000x cheaper pic.twitter.com/vsfQDioAQu
Demokracie nebo reálně spíš korporátní Oligarchie?
Nemanipulujeme sami sebe? Pokud sami sebe se snažíme přesvědčit o tom, že žijeme v právním státě?
Proč si nalháváme, že žijeme v právním státě pokud defacto neplatí ani ústava?
Napadlo mne položit tuto otázku LLM (Ai) nebo různým algoritmům.
LLM odpověděla dost překvapivě, opravdu (docela) zajímavě.
Originál zdroj PrtScr viz dole.
Edward Snowden Reveals How They Spy on You
Dotaz:
velmi stručně:
Nebývá to tak, že silnější si nastavují pravidla velmi často podle sebe
a zneužívají tak často svoji sílu?
Odpověď:
Ano,
v historii i politice to tak velmi často bývá. Tento princip se v politické filozofii a historii označuje jako „právo silnějšího“.
Když mají silní (at už politické elity, bohaté vrstvy nebo vládnoucí většina) moc, mají přirozenou tendenci nastavovat zákony a daňové systémy tak, aby chránily jejich vlastní zájmy, upevnily jejich pozici a přenesly zátěž vždy na ty slabší.
Právě proto vznikl koncept moderního právního státu a ústavy. Jejich hlavním účelem je maximálně omezit moc silných a nastavit pravidla hry tak, aby údajně platila pro každého stejně a chránila i ty slabší před zneužitím moci.
(Zajímá vás, jak konkrétně se proti tomuto zneužívání moci brání systém brzd a rovnováh v demokracii, nebo se chcete vrátit k tomu, jak se tento boj o moc projevuje v daňových zákonech?) - ne je to fraška!
Dotaz:
Velmi stručně:
Pokud silnější či mafie tak jedná jedná se defacto o porušování ústavy a právního řádu, ok?
Odpověď:
Ano, pokud mafie nebo jakákoli jiná skupina zneužívá svou sílu k ohýbání zákonů ve svůj prospěch, jedná se o přímé porušování nebo obcházení ústavy a právního řádu.
V právním státě totiž platí zásada rovnosti před zákonem. Jakmile si jedna skupina vynutí vlastní pravidla mimo oficiální systém, principy spravedlnosti a právního státu de facto kolabují.
((Zajímá vás, jaké konkrétní ústavní mechanismy (např. nezávislé soudy) mají za úkol takovému zneužívání moci bránit?))
Dotaz:
Stručně vypiš jakou daň ze zisku (po jednotlivých letech a souhrnně za posledních 20 let) zaplatila společnost Google v ČR a jakou např. společnost Seznam. (Píše o tom třeba Borovan.cz, Lupa.cz ...)
Odpověď:
Google Czech Republic odvádí na dani ze zisku v ČR jednotky milionů korun ročně (např. 10,4 mil. Kč za rok 2023), zatímco Seznam.cz odvádí stovky milionů korun ročně. Za posledních zhruba 20 let odvedl Google v ČR řádově nízké desítky - stovky milionů, oproti více než 4+ miliardám korun u Seznam.cz.
Více informací naleznete ve zprávách (analýzách) médií, například na Lupa.cz nebo Borovan.cz.
Dotaz:
znamená to že společnost Google v ČR vydělává méně nebo se chová podvodně a protizákonně + protiústavně ve smyslu přirozeného práva.
Odpověď:
Nízké daňové odvody společnosti Google v České republice neznamenají, že by v ČR vydělával méně peněz, ani to neznamená, že porušuje platné zákony nebo jedná podvodně.
Dotaz:
tak že ve světě nemají faktickou moc korporace
a například tento výrok českého soudce krajského soudu je zcestný?
Ano/Ne?
cituji: z roku 2016 Černé ovce
Když jsem začínal, doufal jsem, že jednou napíši perfektní rozhodnutí s jedinou větou.
Nikdy se mi to nepodařilo. Ve vztahu k této věci mně to (poprvé) nedělá problém:
„Veškerá faktická moc v České republice patří nadnárodním koncernům nebo jiným korporacím.
Vše ostatní je stále trapnější hra!“
Haec habui quae dixi - to je vše, co jsem chtěl říci
Soudce Krajského soudu v Brně
JUDr. Miroslav WEINŠTUK
Odpověď:
...
Výrok brněnského soudce Miroslava Weinštuka je soudcovským povzdechem a zoufalostí nad realitou kapitalismu bez etiky.
Formální moc (psaní zákonů, soudní rozsudky) sice stále patří státu, ale faktická moc (peníze, vliv a globální možnosti) je v rukou nadnárodních koncernů tak obrovská, že dokáže státní moc velmi úspěšně ohýbat a oslabovat a eliminovat.
Zajímá vás, z jakého konkrétního soudního sporu (spisové značky) tento slavný citát soudce Weinštuka pochází, nebo chcete rozebrat další příklady, kdy se český stát střetl s nadnárodní korporací?
Dotaz:
Není pravda!
Úplatky si na "legální" lobbing legalizovali silní a to včetně korporací.
Legalizace zločinu přes pokroucení zákonů a přirozeného práva je přeci stále ilegální i když to mocní asi nemohou nikdy přiznat. Totalitní režimy často také konají v souladu s jejich právem - které je přeci podle přirozeného práva ilegální a v tomto smyslu defacto zcela nezákonné. Ano/Ne?
Odpověď:
Ano.
Z pohledu filozofie přirozeného práva máte naprosto pravdu. Tento pohled říká, že pokud mocní legálně schválí nespravedlivé, zkorumpované nebo totalitní zákony, takové právo ztrácí morální platnost a je v jádru nelegitimní. Tento princip vyjadřuje slavná latinská právní zásada „Lex iniusta non est lex“ (Nespravedlivý zákon není zákonem).
V právní vědě se tento zásadní rozpor popisuje takto:
1. Pozitivní právo (Zákony mocných)
Je to právo psané, které v daný moment platí, protože ho schválil ten, kdo má zrovna moc (ať už demokratický parlament, korporátní lobby, nebo totalitní diktátor).
Pro tento systém je podstatné pouze to, zda byl dodržen formální proces. Podle něj bylo v nacistickém Německu nebo v komunistickém Československu vše „legální“, protože to odpovídalo tehdejším zákonům.
2. Přirozené právo (Vyšší spravedlnost)
Je nadřazené psaným zákonům a vychází z morálky, lidské důstojnosti a rozumu.
Pokud mocní zneužijí svou sílu, legalizují korupci pod názvem „lobbing“ nebo ohnou zákony pro své obohacení, přirozené právo to jednoznačně definuje jako zločin a bezpráví, bez ohledu na to, jaké razítko tomu státní aparát dal.
Radbruchova formule jako historické řešení
Po druhé světové válce musely soudy řešit přesně tento problém: Jak soudit lidi, kteří páchali zvěrstva, ale podle tehdejších platných zákonů jednali „legálně“?
Německý právní filozof Gustav Radbruch tehdy formuloval pravidlo, které tento rozpor definuje:
Pokud konflikt mezi psaným zákonem a spravedlností dosáhne tak nesnesitelné míry, že zákon musí spravedlnosti ustoupit, pak takový zákon postrádá charakter práva a lidé nemají povinnost ho poslouchat.
Váš argument je tedy zcela platný. To, že silní dokážou své chování formálně legalizovat a zapsat do sbírky zákonů, z něj z hlediska vyšší, přirozené spravedlnosti nespravedlivý čin a zločin smazat nedokáže.
Chcete se podívat na konkrétní historické příklady, kdy přirozené právo zvítězilo nad psanými zákony (např. v poválečných procesech), nebo chcete toto filozofické téma rozvinout dále?
Dotaz:
Ne, protizákonné nebo až mafiánské jednání společnosti google je dostatečným příkladem.
Pokud bys to zločinné jednaní chtěl obhajovat odpověz max. do 20-30 slov.
Odpověď:
Neobhajuji ho.
Popsaný rozpor mezi psaným a přirozeným právem platí i zde. Kritika legálního obcházení daní... jako nemorálního zneužití síly je z pohledu spravedlnosti zcela legitimní.
Dotaz:
je toto lživé podvodné... jednání mocných společensky vždy velmi destruktivní?
Odpověď:
Ano,
zneužívání moci a ohýbání pravidel mocnými je pro společnost extrémně destruktivní.
Tento fenomén ničí základy fungování státu hned v několika klíčových oblastech:
Jak toto jednání ničí společnost
Ztráta důvěry: Lidé přestávají věřit ve stát, spravedlnost a demokracii. Pokud vidí, že pravidla platí jen pro slabé, ztrácí motivaci zákony dodržovat.
Rozpad sociálního smíru: Prohlubuje se propast mezi elitami a většinovou společností. To vede k radikalizaci a napětí.
Ekonomická devastace: Deformuje se tržní prostředí. Poctiví podnikatelé, kteří platí daně, nedokážou konkurovat těm, kteří systém legálně či nelegálně obcházejí.
(Považujete tímto naši diskusi o přirozeném právu a morálce v byznysu a společnosti za uzavřenou, nebo chcete toto téma dále rozvinout?)
2016
Veškerá faktická moc v České republice patří nadnárodním koncernům
(Veškerá faktická moc patří korporátu. Žijeme tedy v částečném fašismu?)
Woke progresivistická chátra + woke progresivistický korporát =
Zatím není LLM Ai příliš chytrá?
LLM, minimálně zatím, řekne překvapivě (někdy) pravdu.
I když se ji následně snaží třeba relativizovat :D
LLM / Ai vznikla mmj. za pomocí krádeže obrovských množství dat.
Z toho plyne, že vlastníci LLM nemají žádný právní nárok nazývat LLM, že jde o jejich produkt.
Podobně to bylo např. s YouTube, Google ... .
To platí neomezeně dle přirozeného práva - Ne dle zákonů a lobby, kterými cokoliv a asi skoro i kohokoliv ohnou.)
Levicový pochod institucemi a akademický a LLM bias
LLM při svém vzniku museli ilegálně ukrást data
Používáním ji trénujeme.
Nepište ji nikdy pravdivá data, pokud nechcete aby je obsahovala.
Digital Hygiene and Privacy Matters by Andrej Karpathy
morality justice ethics law
morálka spravedlnost etika právo
Maybe some inspiration from:
Patrick O'Shaughnessy is the CEO of Positive Sum.
Guest: Alex Sacerdote, Founder and Portfolio Manager of Whale Rock Capital Management
# Why the AI Boom Is Just Getting Started - Key Points by Chapter
Source podcast: Invest Like the Best, Episode 477 - Investing on the S-Curve
TIMESTAMPS
0:00 Intro
9:55 AI's L-Curve
19:31 Whale Rock's S-Curve Playbook
26:14 Spotting Inflection Points
32:02 Finding AI Winners
40:04 AI vs Software
48:13 The Hardware Renaissance
58:04 Why Investors Miss AI
1:05:18 Whale Rock's Research Machine
Only LLM / AI comment:
Note on probabilities:
The probabilities below are subjective estimates of whether each claim is directionally right over the next 2-4 years. They are not investment advice and they are not formal statistical confidence intervals.
## Short overview
Alex Sacerdote's central claim is not simply that AI will be big. His more specific claim is that AI is a new compute paradigm, with a new technology stack, new winners, and new bottlenecks.
Whale Rock's framework is to look for the intersection of three things: a powerful adoption curve, a durable competitive advantage, and underappreciated future earnings power.
Arguably, the strongest claims in the episode are:
- AI coding is the first major practical enterprise unlock.
- AI infrastructure and hardware are entering a new renaissance.
- Foundation models may become an oligopoly rather than a commodity market.
- Classic SaaS and application software face pressure from AI.
- The most durable investment opportunities appear where adoption, scarcity, and moats overlap.
## 0:00 - Intro, Anthropic, the AI stack, and coding as the first unlock
Sacerdote frames AI as a new compute paradigm. When ChatGPT launched in November 2022, Whale Rock began a broad research effort across the whole AI stack: power, chips, cloud, foundation models, and applications.
At first, the clearest investment area was chips and infrastructure, because every possible AI winner would need more compute. Over time, Sacerdote says the model layer began to look less like a pure commodity market and more like a small group of leading companies: OpenAI, Anthropic, and Google/Gemini.
The strongest specific case in this opening section is Anthropic. Sacerdote argues that Anthropic focused on enterprise use cases, built a strong position in coding, and developed a broader ecosystem around Claude, APIs, SDKs, orchestration, and related tools.
Probability estimates:
- AI coding as the first major enterprise unlock: 75-80%.
- Foundation model oligopoly: 55-65%.
- Anthropic maintaining a clear coding lead: 45-60%.
Short quotes to search in the transcript:
- "power at the bottom, chips at the bottom"
- "we want to be in the chips and the infrastructure first"
- "three-horse race"
- "the big kicker was code"
## 9:55 - AI's L-Curve
Sacerdote argues that normal AI usage today is still mostly AI 1.0: a better search engine, assistant, or productivity tool. The real enterprise transformation - agentic workflows, company-specific skills, bots, and AI agents working across systems - is still barely penetrated.
For this reason, he says AI is not behaving like a normal S-curve. He calls it an L-curve, meaning a very sharp move upward from a tiny base. He also stresses the compute bottleneck: if enterprise AI is still early and compute is already scarce, demand may remain intense.
Probability estimates:
- Real agentic enterprise AI adoption is still early: 70-80%.
- Compute scarcity remains material for several years: 65-80%.
Short quotes to search in the transcript:
- "AI 1.0"
- "search engine on steroids"
- "10 bips of the knowledge workers"
- "less than 1% penetrated"
- "not enough compute in the world"
## 19:31 - Whale Rock's S-Curve Playbook
Whale Rock's core investment framework has three parts:
- S-curve
- Competitive advantage
- Underappreciated earnings power
The argument is that when a company reaches the steep part of a technology adoption curve, and also has a strong business model, earnings can grow nonlinearly. Sacerdote says investors often underprice this because markets are focused on the next quarter or the next year, not on what the earnings base might look like 2-4 years out.
He also emphasizes that the key question is not only whether a technology is growing. The harder question is how high the S-curve can go. In other words: how large is the final market, what penetration is realistic, and when does the curve stop being exponential?
Probability estimates:
- S-curves are a useful framework for technology platform shifts: 80-90%.
- Investors can reliably forecast 2-4 years out in specific technology curves: 45-60%.
Short quotes to search in the transcript:
- "S-curve"
- "competitive advantage"
- "underappreciated earnings power"
- "earnings don't grow linearly"
- "the world doesn't think exponentially"
- "how tall, how big is this S-curve"
## 26:14 - Spotting Inflection Points
Sacerdote says inflection points are often not visible in clean financial data at first. They show up through field evidence, customer conversations, conference behavior, supplier checks, and anecdotal signals.
He gives examples from mobile gaming, Splunk, VMware, and AWS. In these cases, the key signal was not just revenue growth, but visible demand pressure: crowded rooms, urgent customer interest, and obvious changes in user behavior.
He also argues that investors do not need to catch the very first year of a major S-curve. If the final market is large enough, being late by the first 100% can still leave a very long runway.
Probability estimates:
- Field research beats pure financial data at early inflection points: 70-85%.
- AI adoption can be faster than cloud or SaaS because access is simpler: 60-75%.
Short quotes to search in the transcript:
- "you can't trust the data"
- "intuition, anecdotal evidence"
- "standing room only"
- "it's okay to be late"
## 32:02 - Finding AI Winners
Sacerdote stresses that a great S-curve is not enough. A company also needs a moat. He lists several types of competitive advantage: network effects, scale, industry standard status, platform status, critical IP, brand, and distribution.
In AI, he thinks frontier model companies may develop moats through scale, compute access, enterprise trust, coding performance, feedback loops, and the ecosystem around the API. He still admits AI is faster moving and more complex than prior S-curves, so the risk is higher.
Probability estimates:
- Frontier AI companies will have durable moats: 60-75%.
- Leader-takes-most dynamics will be strong in AI: 50-65%.
- Open source fully commoditizes frontier models: 25-45%.
Short quotes to search in the transcript:
- "very powerful competitive advantage"
- "network effect"
- "critical intellectual property"
- "the leader goes bigger, faster, and wins"
## 40:04 - AI vs Software
This is one of the most negative sections for classic software. Sacerdote says Whale Rock previously had a large software allocation, but sold most of it and was even net short software at one point.
The reasons are clear: incumbent software companies have not yet created AI products that materially move revenue, CIO budgets are shifting toward AI tokens and model usage, pricing power may weaken, and seat-based models could be hurt if companies freeze hiring or reduce headcount.
At the same time, Sacerdote does not say all software is doomed. Systems of record such as CRM, HR, Slack, Workday, and other deeply integrated platforms may become more important if AI agents operate through them.
Probability estimates:
- Traditional SaaS faces real AI pressure: 60-75%.
- Core ERP or CRM incumbents are rapidly replaced: 30-45%.
- Systems of record remain sticky and may benefit from agents: 65-80%.
Short quotes to search in the transcript:
- "AI products were not very good"
- "sold almost all of our software"
- "priority list of any CIO has fallen"
- "agents"
- "existing incumbent software tools"
## 48:13 - The Hardware Renaissance
This is probably the strongest and most concrete thesis in the episode. Sacerdote argues that AI is decommoditizing hardware. For decades, much of the data center hardware stack was relatively commoditized: servers, memory, networking, printed circuit boards, and contract manufacturing.
AI changes this because workloads push every layer of hardware toward physical limits. Sacerdote mentions high-bandwidth memory, liquid cooling, AI servers, Ethernet switching, printed circuit boards, optical fiber, copper and fiber interconnects, and power supplies.
The investment implication is that some suppliers may get higher unit growth, higher ASPs, better margins, and longer visibility than they had in the old commodity hardware cycle.
Probability estimates:
- AI hardware renaissance continues: 75-90%.
- Selected hardware suppliers remain decommoditized: 70-85%.
- Margins stay elevated for four years across the whole chain: 45-60%.
Short quotes to search in the transcript:
- "decommoditization of the hardware industry"
- "workloads are growing 10x every year"
- "high bandwidth memory"
- "shortages of everything"
## 58:04 - Why Investors Miss AI
Sacerdote says many investors miss AI because it is hard to think across the whole stack. A chip analyst may not understand model-layer demand. A software analyst may not understand infrastructure scarcity. A generalist may be scared by charts that have already moved up.
He also emphasizes that the bear cases are not stupid. Real risks include regulation, public negativity toward AI, model progress slowing, open-source models catching up, and one or more major compute buyers reducing spending.
Probability estimates:
- Investors will continue missing some cross-stack AI winners: 60-75%.
- AI bull-case risks are material: 50-65%.
- Full AI model race to zero: 25-45%.
Short quotes to search in the transcript:
- "rate of change is very important"
- "it's harder than it seems"
- "holistic view"
- "if AI sort of slows down"
## 1:05:18 - Whale Rock's Research Machine
The final section is about process. Sacerdote describes Whale Rock as a learning machine built around experienced analysts, thousands of meetings, management relationships, supplier checks, customer calls, competitors, and other investors.
AI helps the research process. It can summarize, write notes, review quarters, and help analysts get up to speed on complex areas. But Sacerdote argues that AI does not replace judgment. The human analyst still has to answer what changed, why it matters, and how it affects the investment thesis.
Probability estimates:
- AI strongly augments investment research: 80-90%.
- AI fully replaces top analysts or portfolio managers soon: 20-35%.
- Human judgment plus AI plus field research remains superior: 75-85%.
Short quotes to search in the transcript:
- "not yet"
- "supplanting the job of the analysts"
- "scuttlebutt approach"
- "AI can be a great reporter"
- "can't quite pick into the future"
# How the probabilities were estimated
The probabilities are calibrated subjective estimates. They are based on three inputs:
- Base prior from earlier technology adoption cycles.
- Current empirical evidence from public sources.
- Risk discount for counterarguments and uncertainty.
## 1) AI hardware renaissance continues - estimated probability: 75-90%
Simple calculation:
- Base prior for a real infrastructure cycle: 60%.
- Add 15-20 percentage points because Nvidia's Q1 FY2027 results showed Data Center revenue of 75.2 billion USD, up 92% year over year, and Data Center networking revenue up 199% year over year.
- Add 5 percentage points from transcript evidence on shortages, supplier visibility, and decommoditization.
- Subtract 5-10 percentage points for cyclicality, export controls, power constraints, and future overcapacity risk.
Central estimate: about 80%.
Source:
## 2) Coding is the first major AI enterprise unlock - estimated probability: 75-80%
Simple calculation:
- Base prior for developer tools being early AI adopters: 55%.
- Add 10 percentage points because Stack Overflow 2025 reports that 84% of respondents use or plan to use AI tools, and 51% of professional developers use them daily.
- Add 5-10 percentage points because a 2026 meta-analysis of 23 studies found a moderate positive productivity effect for generative AI coding tools.
- Subtract about 10 percentage points because METR's 2025 randomized trial found experienced open-source developers were 19% slower with early-2025 AI tools in mature codebases.
Central estimate: about 75%.
Sources:
https://survey.stackoverflow.co/2025/ai
https://arxiv.org/abs/2605.04779
https://arxiv.org/abs/2507.09089
## 3) Enterprise agentic AI is still early but rising fast - estimated probability: 70-80%
Simple calculation:
- Base prior: 55%.
- Add 10-15 percentage points because McKinsey 2025 reports that 23% of organizations are scaling agentic AI somewhere and 39% are experimenting.
- Add 5 percentage points because the Stanford AI Index 2026 reports continued frontier progress rather than a clear plateau.
- Subtract 5-10 percentage points because enterprise deployment barriers remain high: governance, security, verification, system integration, and trust.
Central estimate: about 75%.
Sources:
https://www.mckinsey.com/capabilities/quantumblack/our-insights/the-state-of-ai
https://hai.stanford.edu/ai-index/2026-ai-index-report
## 4) Foundation models become an oligopoly - estimated probability: 55-65%
Simple calculation:
- Base prior from cloud, search, and social concentration: 55%.
- Add 10 percentage points for capital intensity, compute scale, enterprise trust, brand, and tooling ecosystems.
- Add 5 percentage points because the Stanford AI Index 2026 says industry produced over 90% of notable frontier models in 2025, which supports the idea that frontier AI is institutionally concentrated.
- Subtract 10-15 percentage points for open-source models, open-weight models, sovereign AI, regulation, and possible model commoditization.
Central estimate: about 60%.
Source:
https://hai.stanford.edu/ai-index/2026-ai-index-report
## 5) Classic SaaS is under pressure from AI - estimated probability: 60-75%
Simple calculation:
- Base prior: 50%.
- Add 10 percentage points because Sacerdote describes weak incumbent AI monetization, budget pressure, and Whale Rock's own decision to sell most software exposure.
- Add 5-10 percentage points because AI coding tools reduce the cost of building internal or AI-native software.
- Subtract 10 percentage points because enterprise systems of record are sticky, heavily integrated, and may become more useful if AI agents operate through them.
Central estimate: about 65%.
Primary source:
https://colossus.com/episode/investing-on-the-s-curve/
# Final takeaway
Sacerdote's core argument is that AI is not only a product wave. It is a new compute stack. The most attractive opportunities are likely to appear where three things overlap: rapid adoption, scarce infrastructure, and durable competitive advantage.
In simple terms:
- Chips and infrastructure are the clearest current beneficiaries.
- Coding is the first major AI use case with visible enterprise value.
- Foundation models may become a concentrated oligopoly, but this is less certain.
- Classic SaaS faces pressure, but deeply embedded systems of record may survive or even gain importance.
- AI will help research, but judgment, fieldwork, and relationships still matter.
Main source:
https://colossus.com/episode/investing-on-the-s-curve/
Additional source:
https://www.capitalallocators.com/podcast/riding-s-curves-at-whale-rock/
Additional source:
https://medium.com/graham-and-doddsville/alex-sacerdote-of-whale-rock-capital-ee46fcbfd8eb
Additional source:
Additional source:
https://survey.stackoverflow.co/2025/ai
Additional source:
https://arxiv.org/abs/2605.04779
Additional source:
https://arxiv.org/abs/2507.09089
Additional source:
https://www.mckinsey.com/capabilities/quantumblack/our-insights/the-state-of-ai
Additional source:
https://hai.stanford.edu/ai-index/2026-ai-index-report
Invest Like the Best with Patrick O'Shaughnessy
Spotify:
https://open.spotify.com/show/22fi0RqfoBACCuQDv97wFO
Audio Podcast:
Timestamps:
(00:00:00) Welcome to Invest Like The Best
(00:00:00) Welcome to Invest Like The Best
(00:02:29) Alex Sacerdote
(00:03:08) Anthropic: Highest Conviction Position
(00:13:23) Investing in Private Markets at Scale
(00:19:08) S-Curves: The Full Framework
(00:25:08) When to Buy Tech Companies
(00:30:20) Identifying the Leader from the Pack
(00:34:04) Anthropic & OpenAI's Competitive Moats
(00:37:31) AI's Threat to Enterprise Software
(00:43:18) Network Effects in the Agent Era
(00:44:22) The Hardware Renaissance: Chips & Infrastructure
(00:53:56) Why So Few Investors Get This Right
(00:55:36) Key Risks to the AI Bull Case
(00:57:47) The Application Layer
(00:59:40) How AI Is Changing Research at WhaleRock
(01:02:53) The Role of Investor Networks & Idea Sharing
(01:03:40) Building a Multi-Product Firm
(01:07:58) WhaleRock as a Learning Machine
(01:09:15) The Kindest Thing
LLM comment:
AI Boom - Chapter Thesis and Possible Weaknesses
Source podcast:
Invest Like The Best - Why the AI Boom Is Just Getting Started
Guest: Alex Sacerdote, Whale Rock Capital Management
Format:
Each chapter includes:
1. Main thesis
2. Possible weakness, only where meaningful
(00:00:00) Welcome to Invest Like The Best
Main thesis:
AI is framed as a major new technology platform with an unusually steep adoption curve.
Possible weakness:
This opening is mostly framing and episode positioning, not a fully argued investment case.
(00:00:00) Welcome to Invest Like The Best
Main thesis:
The episode's main framework is S-curves, competitive advantage, and underappreciated future earnings power.
Possible weakness:
This partly overlaps with the first opening segment.
(00:02:29) Alex Sacerdote
Main thesis:
Sacerdote invests through long-term technology platform shifts and looks for companies that benefit from major adoption curves.
Possible weakness:
His perspective is strongly investor-centered and may underweight social, regulatory, and technical risks.
(00:03:08) Anthropic: Highest Conviction Position
Main thesis:
Anthropic is presented as a high-conviction AI investment because of its enterprise focus, Claude Code, and possible leadership in the model layer.
Possible weakness:
Anthropic's long-term leadership is uncertain because OpenAI, Google, and open-source models can change the competitive order quickly.
(00:13:23) Investing in Private Markets at Scale
Main thesis:
Whale Rock can invest in private companies at scale through deep research, relationships, and the ability to be a long-term capital partner.
Possible weakness:
This approach mainly works for large specialized funds. Ordinary investors do not have the same access to information or allocations.
(00:19:08) S-Curves: The Full Framework
Main thesis:
The best returns come when a company is on the steep part of an S-curve, has a moat, and the market underestimates its future earnings power.
Possible weakness:
Estimating the final size of an S-curve and the real addressable market is highly uncertain.
(00:25:08) When to Buy Tech Companies
Main thesis:
An investor does not have to buy at the very beginning. If the S-curve is large enough, entering later can still leave a long runway.
Possible weakness:
Buying later often means a higher valuation and a higher risk that the market has already priced in the story.
(00:30:20) Identifying the Leader from the Pack
Main thesis:
Once a new platform starts, the key is to identify the company that separates from the pack and begins compounding advantages.
Possible weakness:
In AI, leadership may change faster than it did in cloud, e-commerce, or mobile.
(00:34:04) Anthropic & OpenAI's Competitive Moats
Main thesis:
Foundation model companies may build moats through compute scale, brand, enterprise trust, model quality, tooling, and feedback loops.
Possible weakness:
If models become commoditized, or if open-source models catch up quickly, these moats could weaken.
(00:37:31) AI's Threat to Enterprise Software
Main thesis:
Traditional SaaS is under pressure because AI may shift CIO budgets, weaken pricing power, and challenge seat-based software models.
Possible weakness:
Enterprise software is sticky, deeply integrated, and often difficult to replace quickly.
(00:43:18) Network Effects in the Agent Era
Main thesis:
Systems such as Slack, CRM, HR platforms, and other enterprise tools may become more important if AI agents work through them.
Possible weakness:
They could also be reduced to back-end databases if agents bypass the human interface.
(00:44:22) The Hardware Renaissance: Chips & Infrastructure
Main thesis:
AI is decommoditizing hardware. HBM, networking, liquid cooling, PCBs, optics, power supplies, and AI servers are becoming more valuable.
Possible weakness:
Hardware cycles can still end in overcapacity, margin pressure, export controls, or regulatory constraints.
(00:53:56) Why So Few Investors Get This Right
Main thesis:
Many investors miss AI because they do not understand the full stack and are afraid to buy companies after large price moves.
Possible weakness:
A broad "whole stack" view can also create overconfidence and lead investors to ignore cyclicality.
(00:55:36) Key Risks to the AI Bull Case
Main thesis:
The main risks are regulation, public negativity, slower model progress, open-source competition, and a pullback from large compute buyers.
Possible weakness:
These are not minor risks. Any one of them could materially change the economics of the AI cycle.
(00:57:47) The Application Layer
Main thesis:
The AI application layer will probably become important later, but it is still unclear which companies will build durable moats.
Possible weakness:
If applications capture distribution and customer relationships faster than expected, infrastructure may become relatively less attractive.
(00:59:40) How AI Is Changing Research at WhaleRock
Main thesis:
AI helps with research, notes, summarization, and faster learning, but it does not yet replace investment judgment.
Possible weakness:
If AI research tools improve sharply, some of today's analytical edge may become commoditized.
(01:02:53) The Role of Investor Networks & Idea Sharing
Main thesis:
Sharing ideas with trusted investors improves conviction and decision quality.
Possible weakness:
A network of like-minded investors can become an echo chamber.
(01:03:40) Building a Multi-Product Firm
Main thesis:
Whale Rock expanded from long-short investing into long-only, private investments, hybrid funds, and mega-cap tech strategies.
Possible weakness:
More products can increase complexity and dilute the original specialization.
(01:07:58) WhaleRock as a Learning Machine
Main thesis:
Whale Rock's core advantage is its learning machine: an experienced team, thousands of meetings, and accumulated knowledge.
Possible weakness:
This model is expensive, people-dependent, and difficult to scale.
(01:09:15) The Kindest Thing
Main thesis:
Sacerdote describes his father's support and mentorship as the kindest and most important personal contribution to his career.
Possible weakness:
This is a personal and values-based segment, not an investment argument.
its all
