Frander

322 posts

Frander banner
Frander

Frander

@Frandeeer

AI systems & visual generation.

Katılım Temmuz 2022
61 Takip Edilen130 Takipçiler
Frander
Frander@Frandeeer·
Alex Schultz, Meta CMO and VP of Analytics: The first ad is rarely the winner. It is usually the most expensive way to learn what the winner should have been. That is why AI UGC matters. • 05:09 — why the audience matters before the creative • 14:04 — building a real creative testing roadmap • 17:25 — an AI-made creative case study • 41:25 — using Claude for creative research • 53:15 — the budget and creative-volume question Not because brands need fake influencers. Because brands need more shots at the message before they commit to production. Build one consistent character. Keep the product, face, voice, and style stable. Then test what actually moves attention: the hook, the claim, the story, the objection, the offer, the audience. A traditional shoot turns every new angle into another budget conversation. AI turns it into a creative test. The best use of AI UGC is not replacing the person on camera. It is finding the script worth putting a real person on camera for. The asset is not the video. It is the evidence that one angle worked while nine others did not.
kocer@kocer_eth

x.com/i/article/2069…

English
4
4
20
643
zyxron
zyxron@zyxr0n·
Doyne Farmer, Oxford complexity professor and co-founder of a quant trading firm sold to UBS: A $10,000 trading loss rarely starts with a terrible model. It starts with a model that was right yesterday and nobody told it the market changed. Farmer’s work treats financial markets as an ecology. Every strategy changes the environment. Every edge attracts competition. Every signal eventually gets crowded. That is why an AI swarm is more useful as a system for killing weak theses than predicting the next candle. One agent builds the trade thesis. One tries to break it. One checks whether volatility and liquidity changed. One looks for the correlation nobody noticed. One tracks the catalyst that can invalidate the entire setup. Then the final agent gives the only answer that matters: does this idea still deserve capital? A single model gives you a prediction. A real system gives the prediction enemies. The edge is not making AI more confident. It is making AI notice when confidence is no longer justified.
zostaff@zostaff

x.com/i/article/2075…

English
3
1
10
228
Frander
Frander@Frandeeer·
Joseph Weizenbaum, MIT professor and creator of ELIZA: He built one of the first chatbots to show how easily a machine could imitate understanding. People trusted it anyway. 60 years later, GPT-4.5 was picked as human 73% of the time in a controlled Turing test when it was given a human persona. Without the persona, it fell to 36%. Same model. A completely different social signal. That is the part most AI discussions miss. The breakthrough is not only that models can reason. It is that they can now produce the cues people mistake for a mind: hesitation, warmth, memory, uncertainty, imperfection, intent. Weizenbaum called attention to the original problem decades ago. Humans do not need a machine to understand them before they start feeling understood. The next AI risk is not only wrong answers. It is false trust at scale.
Frander@Frandeeer

Alan turing gave us the test. 75 years later, gpt-4.5 passed it with a 73% win rate when it was given a human persona. The important detail is not that people got fooled. people have always been easy to fool. the important detail is that a small change in the prompt moved the model from obviously artificial to more believable than a real human. that is the new moat. not a smarter answer. a system that knows when to sound certain, when to hesitate, when to be imperfect, and when to feel human. the first thing AI automated was not intelligence. it was the feeling that intelligence is present.

English
1
8
32
2.1K
Frander
Frander@Frandeeer·
@kocer_eth Fable 5 is a legend; I don't know how I'd live without it
English
0
0
0
33
kocer
kocer@kocer_eth·
WITH FABLE 5, ONE EVENING TURNED A ROCKET LEAGUE IDEA INTO A PLAYABLE 3D LOOP: CAR, BALL, GOALS, SCORE, TIMER. not a cinematic mockup not a pitch deck not a fake “game concept” trailer The video shows a 3D car-soccer prototype called Nitro League. Blue car at kickoff. Ball physics. Neon goals. Blue vs orange scoring. A countdown timer. A GOAL screen when orange scores. The important part is not “AI recreated Rocket League.” The important part is the shape of the work. Before, the blank page in game dev was brutal: set up the arena make the car move make the ball collide detect goals track teams show score keep time make the loop playable enough to judge Now the first ugly version can appear fast enough that the real design work starts the same evening. That changes what prototyping means. You are not waiting all weekend just to find out if the core loop exists. You can get the loop on screen, touch it, break it, and start asking better questions. Does the car feel heavy enough? Is the camera readable? Does the ball bounce right? Is recovery after a hit annoying? Does scoring feel satisfying? Is there actually any fun here? Fable 5 does not remove that taste layer. It compresses the distance between “I have an idea” and “I can test the playable version.” That is the useful shift. AI game dev is not finished-game craft on demand yet. It is the destruction of the blank-canvas phase.
Asteri@Asteri_eth

x.com/i/article/2077…

English
9
3
26
1.8K
kocer
kocer@kocer_eth·
@Frandeeer the psychology of prompt tuning is wild
English
1
0
1
54
Frander retweetledi
Frander
Frander@Frandeeer·
Alan turing gave us the test. 75 years later, gpt-4.5 passed it with a 73% win rate when it was given a human persona. The important detail is not that people got fooled. people have always been easy to fool. the important detail is that a small change in the prompt moved the model from obviously artificial to more believable than a real human. that is the new moat. not a smarter answer. a system that knows when to sound certain, when to hesitate, when to be imperfect, and when to feel human. the first thing AI automated was not intelligence. it was the feeling that intelligence is present.
DiKrass -X-@Di_Krass_

x.com/i/article/2075…

English
10
5
61
4.5K
Frander
Frander@Frandeeer·
GPT-5.6 Sol costs $30 per million output tokens. One silent model mistake can cost 1,000x more. OpenAI just split GPT-5.6 into three tiers: Luna: $1 input / $6 output. Terra: $2.50 input / $15 output. Sol: $5 input / $30 output. Most teams will read that as a pricing table. They should read it as a permissions table. A low-cost model can: classify inbound email, clean CRM records, extract invoice fields, draft customer replies, prepare a first-pass report. It should not silently: approve a payment, migrate a database, change production access, or summarize a legal risk without source-level proof. The right routing rule is simple: cheap model for repetition. balanced model for daily execution. frontier model for decisions where failure hides below the surface. Then add: proof, a human gate, a retry cap, and a stop condition. OpenAI made a stronger model family. The real upgrade is the mental model: not “which model is best?” but: which model deserves the right to act here? Prompt ↓
Diam@diamai_

x.com/i/article/2075…

English
7
1
19
720
Frander
Frander@Frandeeer·
You are an AI model-routing architect. For each workflow below, decide: 1. Which model tier should handle it: - cheap worker - balanced operator - frontier reasoner 2. What permissions the model may receive 3. What proof it must produce before acting 4. Whether a human approval gate is required 5. The retry cap and stop condition Workflows: [paste your workflows here] Output a table: workflow | model tier | permissions | required proof | human gate | retry cap | stop condition
English
0
0
0
37
zyxron
zyxron@zyxr0n·
$1M revenue - $30K profit - one equation worth $10K. Most companies think AI creates value by doing work faster. The bigger win is making work wait less. John Little proved the math in 1961: work in progress = throughput × waiting time. Every unread email, unmatched invoice, unapproved quote, or unassigned job is not just “admin.” It is money sitting in a queue. A business at 3% margins has almost no room for waiting. Remove $10K of coordination cost: $1M revenue $30K profit becomes $40K profit. Same customers. Same prices. No magical growth strategy. The useful AI agent is a queue monitor. It finds the missing information, moves the task forward, and sends a human only the exception. The future is not AI replacing employees. It is AI removing the time between one decision and the next. Prompt ↓
zyxron@zyxr0n

$1M REVENUE. $30K PROFIT. REMOVE $10K IN COORDINATION COSTS AND KEEP $40K AI’S BIGGEST WINNERS MAY BE THE COMPANIES NOBODY CALLS AI COMPANIES. A software company at 30% margins uses AI and becomes slightly more efficient. A logistics, manufacturing, staffing, or field-service company at 3% margins can change its entire earnings profile. The opportunity is not replacing the person doing the work. It is removing the coordination tax around them: dispatching, scheduling, routing changes, customer updates, invoice matching, approvals, claims, exception handling, back-office reconciliation. This is the invisible layer draining low-margin businesses. The math is simple: $100 revenue. $97 operating costs. $3 profit. Remove $1 of coordination cost and profit becomes $4. Same customers. Same prices. 33% more profit. But this will not happen by giving every employee another chatbot. The agent has to operate inside the tools where work already lives: email, PDFs, spreadsheets, NetSuite, inboxes, approval systems. It reads the invoice. Matches it to the purchase order. Flags the exception. Prepares the approval. Routes only the judgment call to a human. Then learns from the decision. This is not AI as another tool employees must remember to use. This is AI as infrastructure the P&L cannot ignore. Prompt ↓

English
3
1
8
132
Frander
Frander@Frandeeer·
Bill Gates reportedly put a $500 million short on Tesla. Then he messaged Elon Musk asking to discuss climate philanthropy. That conversation went about as well as you’d expect. Musk asked: “Do you still have a half-billion-dollar short position against Tesla?” Gates replied that he had not closed it. In the same message, he said he wanted to talk about philanthropic opportunities. Musk’s response was blunt: He said he could not take Gates’ climate work seriously while Gates was betting heavily against Tesla — the company Musk believed was doing more than almost anyone to accelerate the shift to electric vehicles. That is what made the situation so explosive. A short position becomes more profitable as a stock falls. In the most extreme case, the biggest payoff comes if the company collapses. So one of the world’s most prominent climate philanthropists was simultaneously positioned to benefit from Tesla failing. Gates later described Musk’s reaction like this: “Once he heard I’d shorted the stock, he was super mean to me. But he’s super mean to so many people, so you can’t take it too personally.” The position reportedly remained open. As Tesla’s stock surged, that $500 million bet was later estimated to have turned into a loss of roughly $1.5–2 billion. Musk then publicly mocked Gates after the private messages leaked, posting a meme with the caption: “In case u need to lose a boner fast.” He later added another warning: “If Gates hasn’t fully closed out the crazy short position he has held against Tesla for ~8 years, he had better do so soon.” Two billionaires. One enormous short position. Private texts leaked to the public. And a reported multibillion-dollar loss built around a personal feud.
localminima@localminimaa

x.com/i/article/2075…

English
1
2
18
2.1K
Frander
Frander@Frandeeer·
@Rossst_03 It's really interesting to hear from people with that kind of capital and experience I'll keep this in mind
English
0
0
1
131
Rossst.03
Rossst.03@Rossst_03·
Ray Dalio, billionaire founder of Bridgewater ($150B+): "In 1982 I was totally certain the economy would collapse. I bet everything on it, and I was so wrong I nearly went bankrupt. That's when I learned confidence is not skill." the article above is about telling a real edge from luck. dalio built the biggest hedge fund on earth only after learning the hardest way that being certain and being right are not the same thing. he had every reason to feel like a genius. he'd been right before, his conviction was total. then the market wiped him out, and all that confidence turned out to be worth exactly nothing. it didn't care how sure he was. so he made one change that saved his career: he stopped asking "am I right" and started asking "how do I know I'm right." he treated his own certainty as the enemy. that shift is the entire gap between the fool he was in 1982 and the investor he became. he tells the whole story in this free TED talk. here is the trap. a trader on a winning streak feels the exact certainty dalio felt in 1982. it feels like skill. usually it is just the market not having punished you yet. the edge is doubting yourself hardest when you feel surest, stress-testing the idea you love most, and sizing so that being wrong can never end you. the math is free. treating your own confidence as a warning sign is the discipline almost no one has.
Rossst.03@Rossst_03

x.com/i/article/2075…

English
10
16
113
16.8K
Frander retweetledi
Frander
Frander@Frandeeer·
David Bailey, mathematician at Lawrence Berkeley National Lab: He gave the problem its name. Backtest overfitting. And he proved what most quants never say out loud. Test only a handful of strategy variants, and a beautiful Sharpe is almost guaranteed. Even when the true edge is exactly zero. He published it in the Notices of the American Mathematical Society, under a title most funds would rather ignore: Pseudo-Mathematics and Financial Charlatanism. This free breakdown turns his math into runnable numpy/scipy code. It opens with one experiment. Generate 200 strategies that are pure noise. Zero true edge, by construction. Then do what every naive swarm does: pick the highest backtested Sharpe. The winner comes back near 1.6 annualized. The number a discretionary trader would quit their job for. Clean equity curve. Passes every eyeball test. Would get funded on the backtest alone. It is nothing. You selected the luckiest coin in a bag of fair ones. A swarm testing 100 hypotheses a night does not escape this. It runs into it 100 times faster. Bailey's two tools are built to catch exactly that: The Deflated Sharpe Ratio raises the bar a winner must clear as the trial count climbs. The Probability of Backtest Overfitting asks whether the selection procedure generalizes, or just fits noise. Both depend on one number. The honest trial count. Every swept parameter, every discarded variant, every abandoned idea. Each one increments N, and N feeds the correction. Report only the winner without carrying N forward, and you have disabled the one gate that would have caught it. This is not proof that no swarm can find alpha. It is proof that throughput without honest counting manufactures the illusion of it. Do not ask how many strategies your swarm can test. Ask how many you counted honestly.
localminima@localminimaa

x.com/i/article/2075…

English
2
8
21
2.7K
Frander
Frander@Frandeeer·
@VoltexGar Bro, this is a great post—it's really well written. Thanks, I'll save it.
English
0
0
0
23
Voltex
Voltex@VoltexGar·
Professor David Hand, Imperial College statistician: "A one-in-a-billion trading streak isn't a miracle. It's a mathematical certainty, and I wrote the book on why the incredible happens every single day." the article above is about telling a real edge from luck. david hand wrote a whole book on why the incredible is actually ordinary, "The Improbability Principle." his core law is brutal: with enough opportunities, extremely unlikely events aren't just possible, they're guaranteed. run millions of traders and one posts a flawless five-year record by luck alone. it had to happen to someone, and he can compute how many. then add what he calls the law of selection, we only ever hear about that one winner, so the miracle looks even bigger than it was. so a track record that feels impossible to explain by chance usually isn't. "impossible to explain by chance" is exactly what chance manufactures at scale. your streak feels like destiny only because you can't see the thousands who ran the same play and vanished. he taught this at imperial for decades, and this lecture is free. same point as my article: an amazing run is a question, not proof. here is the edge. refuse to be impressed by your own results until you've asked how easily luck alone could have produced them. the math is free. that discipline is the rarest thing in the game.
Rossst.03@Rossst_03

x.com/i/article/2075…

English
7
1
23
2.9K

Keşfet

@zyxr0n @kocer_eth @Di_Krass_ @biektive @Rossst_03 @elonmusk @BarackObama @taylorswift13