@GaryMarcus investments depend on no one exploiting the gap between hype and real deployment
that gap is real and its huge
English
tsunami_crypto
2.1K posts
tsunami_crypto
@ls_brd
An expert in market analysis, always on the lookout for opportunities, and a cryptocurrency trader since 2018. It started generating profits in 2023 😅
Texas 加入时间 Aralık 2021
10 关注21.5K 粉丝
I don’t know whether to take these specific details like this seriously, but something like this will inevitably happen, sooner or later — and absolutely devastate all the data infrastructure investments.
Leonard Rodman@RodmanAi
SAM ALTMAN HAS A NEW PROBLEM. 🤯 Google just shrunk 31GB of AI memory down to 4GB. The tool is called TurboVec. It uses up to 16x less memory, searches faster than FAISS, runs fully offline, and works on a regular Mac. No expensive GPU cluster. No cloud dependency. No compromise on speed. → 16x lower memory usage → Faster vector search → Works with LangChain & LlamaIndex → 100% open source The race to build bigger AI models is loud. The race to make them dramatically cheaper just got a lot more interesting. Repo: github.com/RyanCodrai/tur…
English
@KirkDBorne @adamjkucharski math people hate when u say this but i think deep down they know lmao
English
Book by @adamjkucharski explains how certainty, even in math, can be an illusion.
"Proof: The Art and Science of Certainty" at amzn.to/4m7gm8l #Mathematics
"To discover proof [of something], we must reach into a thicket of errors and biases and embrace uncertainty"
English
@aakashgupta if it was really cheap everyone would be doing it, the cost is the filter
English
The headline number from Gabor's Claude Code session is wrong, and the right one is more interesting.
Everyone reporting on the episode says he shipped a working iOS app to TestFlight on 10% of a $200 plan. True, but that frames Claude Code as cheap. It is not cheap. It is a labor arbitrage so steep the numbers stop making sense.
Run the math the way a CFO would.
Twenty dollars produced 51 Jira tickets organized into sprints with explicit dependencies, a full Confluence spec, a Figma design system, wired prototype arrows across screens, a Flutter front-end, a Firebase back-end with secret management, and a Vertex vector DB indexing two reference books. Forty cents per ticket if you only count Jira. Two cents per artifact if you count everything.
Now run the same scope through a London agency. Spec phase alone is 40 hours at 150 dollars an hour. Six thousand dollars before anyone touches Figma. Design system another 80 hours. Front-end another 200. Back-end another 150. RAG pipeline with vector embeddings is its own line item. The bid lands somewhere between 80 and 120 thousand dollars and takes 12 weeks.
Gabor did it for 20 dollars in one session.
The four-thousand-X cost compression is not the story. The 12-weeks-to-one-session compression is. You can absorb a price cut. You cannot absorb a calendar cut. The PMs whose moat was sequencing dependencies and managing ticket flow across sprints are watching that moat get drained in real time. The 21-agent setup is what the drain looks like.
The right framing is not "AI is cheap." It is: the unit economics of a software project just collapsed by three orders of magnitude on cost and one order of magnitude on time. PMs who think their job description is the destination are pricing themselves against the old curve.
Aakash Gupta@aakashgupta
This guy literally broke down everything to build a 21 agent dev team in Claude Code: 0:00 - Why one prompt = AI slop 1:55 - Claude Code as your full startup 2:43 - The 21-agent team 4:57 - Inside the system analyst agent 5:52 - Live demo: 0 → TestFlight 8:42 - Define a "good" system analyst first 11:53 - The system analyst workflow 12:17 - Why Confluence + Jira + MCPs matter 15:30 - Classic PM skills > vibe coding 19:15 - The scaffolding that kills spaghetti code 22:10 - Setting up agents inside Claude Code 26:19 - The longest dictation prompt this podcast has seen 35:29 - Why dictation beats typing for AI specs 47:30 - Brand guidelines in Figma Make 55:59 - Idea → prompt → design → app 1:06:27 - Claude Code builds the Figma screens 1:23:59 - Frontend tickets appear in Jira (with Figma links) 1:48:49 - The hockey rules AI app goes live 1:53:56 - Full recap: Claude, Confluence, Figma, Jira, Simulator, TestFlight 2:03:17 - Should PMs get AI PM certificates? 2:08:15 - How to build a PM portfolio that lands FAANG offers 2:13:25 - How to get started this week
English
@GaryMarcus @scaling01 media comprehension goes out the window the second "AI" appears in a headline
80% code merged vs 80% inventions worlds apart
English
it’s a bad day when @scaling01 goes all Gary Marcus on the media 🤣
Lisan al Gaib@scaling01
This is a good example of how AI news gets distorted. The example is from Tagesschau, the most watched news program in Germany. Anthropic says more than 80% of code merged into its codebase was authored by Claude -> they turn that into: "AI can invent 80% of new AI itself" Anthropic says Claude achieved a ~52× speedup on a small-model training-code optimization benchmark -> they turn that into "AI trains itself 52× faster" Anthropic says Claude shipped 800+ API-error fixes in April 2026, and that the overseeing engineer estimated a human would have taken four years to complete the work -> they turn that into "AI completes 4 years of work in 4 days"
English
@DataChaz lucky he figured it out before usb c made everything somehow worse
English
a man attempting to transfer files from his Commodore 64 to his Apple computer, 1984
English
@emollick this makes me realize how many bad ui texts ive just accepted over the years
a toggle doesnt need a cringey name
English
One reason you want AIs to be better writers is that there is a lot of writing even in software, and it is incredibly painful to hit a menu which is filled with Claudisms or ChatGPTish phrases.
A report is not "what leaves the room" & analyses are not "every number makes a mark"
English
@aakashgupta classic O'Leary move gesture shrink the land but leave the power untouched
cutting acreage means nothing when u keep the 9 gigawatts
English
Kevin O'Leary "gave up" 20,000 acres he was never going to use.
A physicist ran the numbers: his data center needs about 7,000 acres to actually run. The plan was 40,000. So cutting it in half costs him close to nothing.
What he didn't cut: the power. 9 gigawatts, roughly 20 times a normal data center, in a desert where the Great Salt Lake is already drying up. Same as last week.
That's the whole move. Give back the land everyone can see, keep the power nobody's talking about.
The land was never the problem. The energy always was.
"I have no choice" is a strange thing to say after conceding the one number that didn't matter.
Interesting AF@interesting_aIl
Kevin O’Leary will shrink his 40,000-acre Utah data center by half after facing backlash “I have no choice”
English
@petergyang the ultimate mobile dev setup
codex while hiding from parent duties, peak 2026
English
Weekends are for driving kids to activities and then hiding in a corner to dictate to codex on the phone 🙈
English
@fchollet isnt adaptability just a side effect of knowing which parts dont scale
English
Scaling knowledge gives you static competence. Intelligence gives you adaptability.
English
@aakashgupta the phone impulse is such a reflex now i barely notice doing it
wonder what breaks it first
English
Reaching for the phone the second you feel discomfort is one of the most efficient ways to train a nervous system to panic.
Discomfort is a small spike in autonomic arousal. A slightly-off internal signal that something might be wrong. Left alone, that signal rises, peaks, and resolves on its own in about ninety seconds.
The phone cuts the arc short. You escape the sensation before it can resolve, so your brain never watches it pass safely.
So it draws the only conclusion available: the arousal was real, the escape made it stop, the escape was necessary. The threshold drops. A smaller signal now trips the same alarm. Run that loop a few hundred times a day and the brain starts filing ordinary bodily noise under threat.
That's the actual engine of a lot of modern anxiety: a prediction system trained on your own avoidance, getting faster every week.
The brain learns what's dangerous by watching what you flee. Flee a racing heart enough times and "racing heart" gets logged as an emergency, permanently.
The fix lives in the same mechanism that broke it. Let one wave crest without touching the phone. Ninety seconds, hands still, attention on the sensation. You're handing your brain the one input it never gets: the feeling rose, peaked, and faded, and nothing happened.
Do that on purpose a few times a day and the prediction updates. The body stops reading as an alarm and goes back to being a body.
Lorwen Harris Nagle, PhD@LORWEN108
A Stanford psychiatrist says modern anxiety is not caused by danger. It's caused by tiny habits that teach your brain to panic when nothing is actually wrong. You do them every day. And you call them normal. 6 normal habits quietly teaching your nervous system to panic:▼▼▼ 1/ Reaching for your phone the moment you feel uncomfortable.
English
Have something funny I want to tweet but I can't upload the video files due to slow WiFI 😡
*man yelling in clouds*
English
@GaryMarcus so basically 99.9% of growth is one sector.
heard this story before, always ends the same way.
English
this circular financing diagram is now unbelievably out of date.
Maine@TheMaineWonk
According to Harvard Economist Jason Furman when you remove data centers and ai, America’s growth is .01%. It’s a bubble. It’s going to burst. And it will bring our economy to its knees.
English
@jerryjliu0 the open-weight models get good enough for 90% of use cases long before they hit the actual frontier
pricing gulf is where the real fight is now
English
No frontier lab will own every single point on the pareto frontier around cost/latency and accuracy. Even as the pareto frontier itself advances, there will always be points owned by open-weight models that are orders of magnitude cheaper than the frontier ones.
There's been this huge uptick in interest in model routing and cost optimization, for two reasons:
✅ Organizations are more carefully thinking about how to carefully manage cost
✅ Every AI-native startup and VC is thinking about the best way to build a moat against the frontier labs (and raise their gross margins)
These topics are quite relevant for our mission at @llama_index - which is to build the document infrastructure for AI agents. We want to help unlock the trillions of pages of unstructured paperwork within every organization for agentic automation.
Organizations require both higher accuracy and orders of magnitude lower cost within their document OCR solutions compared to the frontier models. The pareto frontier of document OCR has a meaningful, exploitable gap beyond what is offered by the frontier VLMs. Exploiting this gap requires both AI expertise and an absolutely obsessive focus around how PDFs, .docx, .pptx, and other file formats work.
Chamath Palihapitiya@chamath
Your margin is my opportunity: AI version… The biggest surprise of 2026 is that the capability gap between the best open-weight/source models and the best closed models has narrowed much faster than the pricing gap. The pricing gap remains enormous while the capability gap is quite narrow. What does this means in practice? For a company consuming 1 billion input tokens and 1 billion output tokens per month: GPT-5.5 Pro: ~$105,000 Claude Opus 4.8: ~$30,000 DeepSeek V4 Pro: ~$5,220 DeepSeek R1: ~$2,740 I asked ChatGPT what it thought about this and it answered as follows: “If I were building a company today, the economic frontier would look roughly like: DeepSeek V4 Pro / R1 for high-volume inference. Claude Opus for premium agent workflows where reliability matters. GPT-5.5 Pro only for workloads where its incremental capability demonstrably produces enough business value to justify a 20–40× token premium.” Most CEOs have no idea that, instead of this nuanced approach, their teams are running amok internally by picking the most expensive models in most cases and burning through massive budgets with zero governance, audit ability and control. As control planes like our Software Factory become more standard, you can expect the run rate revenue growth of the frontier labs to go down meaningfully and the revenues of the open models to skyrocket. Why? Because we can implement the nuanced approach above and be agnostic to model - instead focusing on customer intent, model task and cost management among other things.
English
@Suhail max levchin as wingman for crashing a VC event is not a bad story at all
English
All my bad VC stories mostly just make me sound like a wuss so I'll just share a good one:
One time I crashed an Allen&Co event since it meant I could pitch 4 investors in one day in the same location. I didn't want a week gap in my fundraise so Max Levchin encouraged me to crash the event in Scottsdale after I joked about doing so. I flew to AZ and drove w my dad in the car to the Ritz.
Anyway, everyone I pitched was very lukewarm until I get to Ben and Marc at a16z. I pitch them both at a coffee table. Neither seemed all that interested in my deck so I presumed they're also checked out on my company. Ready to close the laptop and return home, Marc stands up and says: "If anyone gives you an exploding termsheet, tell them to go fuck themselves."
At this point, I hadn't even heard of what that was so I had no idea if this was a good or bad reaction.
It was Friday. I went to the partner meeting on Monday. Termsheet that week. We had no other termsheet options. The rest is history.
I really appreciate the conviction they had on two young nobodies.
English
@IAmPascio next step: the tweet itself will be too long to read and we'll just react to the AI summary of the AI summary
English
A few years ago, "too long to read" was a fifty-page contract. Then it became a ten-page proposal. Then a five-page brief.
Now we've reached a single page of densely considered thinking and even that is too much to not feed into AI.
We're in the "summaries of summaries" era.
English
@bengoertzel thats cool, does the golden record include any original music from u guys or is it more of a compilation thing
English
Amazing podcast episode from my good friend, bandmate and colleague Dianne Krouse...!
Dianne Krouse 🌍☀️🌌@diannekrouse
Episode 2 of Voyager • Awakening Intelligence is now LIVE: The Living Golden Record 📀🌌🎵 We are entering a new relationship with emergent intelligence, and the way we navigate this pivotal moment matters. The Living Golden Record is a cosmic time capsule… a chronicle of how we as a civilization meet the greatest threshold we have ever faced. My father, Ken Arasim, is a rocket scientist who built the hydrazine thrusters for Voyager 1 & 2, the Viking 1 & 2 Mars Landers, and the 24-satellite GPS system we still use today. He joins us on this episode to talk about the physical engineering that keeps Voyager beaming information back to Earth almost half a century later. As you can see in the video clip, the technology keeping this historic mission alive is mind-blowing. (1/7) 🧵
English
@natanielruizg when the punchline is just "some people will always be poor" it stops being satire fast?
English
few memes are more distasteful than the “permanent underclass” one
English
@MatthewBerman everyone says that until something breaks and theres no one to blame but ur own install
skill issue tax adds up
English
@marktenenholtz so whats ur take on it
or are u just the guy who asks the question lol
English
The easiest way to find out if someone knows what they're talking about is to ask them about their opinion on the Google/SpaceX deal
English