Replying to @GoogleDeepMind
@GoogleDeepMind shipped Gemini 4 Neon this week, a mid-tier model at $3 per million input tokens and $15 per million output. My honest reactions:
1/ The middle tier is where the real fight is now. Most production traffic never needed the top model. It needed a good one cheap enough to run all day.
2/ Price per token is the wrong number to compare. A cheaper model that needs three tries can cost more than a pricier one that gets it right once. Cost per finished task is the number that matters.
3/ Launch benchmarks are the lab grading its own homework. Give it two weeks of people running real workloads before you move anything.
4/ The big labs all ship a ladder of tiers now, so for a lot of teams switching labs is a config change. That's good for buyers and bad for anyone counting on loyalty.
5/ The business angle: if your product is built around one model, you built on a price you don't control. Route by task and keep the option to swap.
2
1
327
We book meetings for B2B companies on a pay-per-meeting basis.
You only pay for qualified prospects who actually show up.
If you're interested, book a call with us here:
introslot.com/revenueflow/ca…
95
Satya Nadella was asked if Elon Musk is right that AI compute is going to space. His answer: "if he has space capacity that I can lease, I would lease it."
That's the smartest answer on the table. A lease leaves Musk carrying the rockets, the cooling and the risk, and @Microsoft only pays if the thing works.
@andrewrsorkin asked @satyanadella at @dealbook's summit this week whether it all happens within three years. "I mean, he is the expert on it."
He added that the folks at @awscloud say it isn't happening for a very long time.
@elonmusk replied to a post of the quote with one word: "Cool".
From @dealbook.
2
1
638
We book meetings for B2B companies on a pay-per-meeting basis.
You only pay for qualified prospects who actually show up.
If you're interested, book a call with us here:
introslot.com/revenueflow/ca…
117
Dario Amodei opened Anthropic's first day as a public company with a story about his father.
I think the story was the right one, and the honest part came later in the speech. Discovery was never the only slow step in medicine, so a century of progress in a decade depends on AI speeding up trials and approvals. Reading papers faster is the easy half.
@DarioAmodei, at the @NasdaqExchange opening bell on Thursday: "Daniela and I lost our father to a disease that just a few years later became curable. I started my career in biology hoping to help future cures come sooner."
His bet: "I believe sufficiently powerful AI working in concert with humans could compress a century of scientific and medical progress into a decade."
His own example of how slow it is today was CRISPR. The repeating DNA pattern behind it was first reported in 1987. The first CRISPR treatment was approved in the UK in November 2023, 36 years later. A model that reads every paper overnight shortens the front of a timeline like that. The back of it is patients, trial sites and regulators.
He said as much himself: "At this point all we have are early results. Experiments and clinical trials will still take time and although they involve slow and ponderous processes I believe these processes can be substantially accelerated by AI."
That last sentence is the whole bet. If AI only speeds up the lab, the wait for a cure barely moves. If it speeds up the trial, his closing line comes true: "We hope that in the future, fewer families miss a cure by a few years."
His sister and co-founder, @DanielaAmodei, is president of @AnthropicAI.
From @Nasdaq.
We book meetings for B2B companies on a pay-per-meeting basis.
You only pay for qualified prospects who actually show up.
If you're interested, book a call with us here:
introslot.com/revenueflow/ca…
123
Strauss Zelnick, the CEO behind GTA 6, says AI has not cut Take-Two's back office. His best example of what it does do is a pie chart.
I think that pie chart explains why so many CEOs say AI has not cut jobs yet. The first work it takes is work nobody would have hired for, and headcount only moves once it does work someone would miss.
On launch day, @strausszelnick told @tbpn's @johncoogan and @jordihays: "for those who think you're going to vastly reduce your head count overnight, that clearly is not happening."
Where it is paying off: "it's quicker to make a pie chart than it used to be. Pie charts don't create any value whatsoever."
And on cost: "we are spending money on tokens, which is something we have to think about."
From @tbpn with @strausszelnick.
We book meetings for B2B companies on a pay-per-meeting basis.
You only pay for qualified prospects who actually show up.
If you're interested, book a call with us here:
introslot.com/revenueflow/ca…
1
99
Vibe-coded internal tools are the new shadow IT. A few observations:
1. Building the dashboard was never the hard part. Keeping it right after the person who prompted it changes teams is.
2. Most of these apps live in one chat, on one laptop, behind one person's login.
3. IT won't ban them. It will ask who owns them, and nobody will have an answer.
4. The tools that win this space will look boring: hosting, permissions, an audit log, a named owner.
5. The real cost of a vibe-coded app shows up the day it is wrong and someone already made a decision with it.
We book meetings for B2B companies on a pay-per-meeting basis.
You only pay for qualified prospects who actually show up.
If you're interested, book a call with us here:
introslot.com/revenueflow/ca…
99
Nick Bostrom took the title of @ESYudkowsky and @So8res's book, "If Anyone Builds It, Everyone Dies," and turned it around: "if nobody builds it, everyone dies."
He's right that waiting has a body count too, and it is the number the pause debate keeps leaving out. A pause should have to show the safety it buys, the same way a launch has to show it is safe.
His conclusion, in the clip: "So you don't want to delay a single day longer than you really need to."
From @a16z.
We book meetings for B2B companies on a pay-per-meeting basis.
You only pay for qualified prospects who actually show up.
If you're interested, book a call with us here:
introslot.com/revenueflow/ca…
139
Fei-Fei Li says universities are being priced out of AI.
She is right, and I think it's worse than a funding gap: if the only people who can afford to test frontier models are the people selling them, nobody independent gets to check the claims.
@Edgecliffe asked her if universities are keeping up:
"I don't think universities and public sectors are keeping up because we're priced out."
@drfeifei co-founded @theworldlabs and was founding director of @StanfordHAI, so she sees both sides of the bill:
"we need tokens and compute to discover the cure for cancer, to teach the next generation, to also hold the open society responsible as a key voice and key player in understanding and benchmarking this technology."
Benchmarking is the word I'd underline.
From @semafor with @drfeifei.
1
687
We book meetings for B2B companies on a pay-per-meeting basis.
You only pay for qualified prospects who actually show up.
If you're interested, book a call with us here:
introslot.com/revenueflow/ca…
123
Andy Jassy says @awscloud took three years to reach a $57 million run rate, and that AI revenue is over $25 billion three years into this wave.
That ratio is carrying the whole case for Amazon's $220 billion of capex, and I'd watch it more closely than the capex. In April he put the AI number at over $15 billion.
@ajassy, asked by @Kr00ney what justifies the spending:
"So about 440 times faster than the way we grew in AWS when we thought we were growing fast."
From @semafor.
1
771
We book meetings for B2B companies on a pay-per-meeting basis.
You only pay for qualified prospects who actually show up.
If you're interested, book a call with us here:
introslot.com/revenueflow/ca…
126
Unpopular opinion: most cold email deliverability problems are list problems.
When a campaign lands in spam, the reflex is to buy more inboxes, warm them longer and rotate domains.
But inbox providers pay close attention to how people treat your mail. Send to the wrong people and they ignore it, delete it or mark it as spam, and that is the signal that sinks you.
Send to people who obviously need the thing and they open it, read it and sometimes reply. Same copy, same setup, very different placement.
So before touching the infrastructure, look at the list:
1. Would every person on it recognise the problem from one line?
2. Does their role actually own that problem?
3. Did something happen at their company that makes now the time?
Fix the list first. The infrastructure is usually fine.
We book meetings for B2B companies on a pay-per-meeting basis.
You only pay for qualified prospects who actually show up.
If you're interested, book a call with us here:
introslot.com/revenueflow/ca…
130
Brian Armstrong runs seven or eight AI agents at once, plus an AI chief of staff that oversees them. The first thing to hit its limit was him.
I think that is the real lesson in this clip: the agents scale fine, and the person reading their output does not. Whoever designs the review queue will matter more than whoever picks the model.
@brian_armstrong, the co-founder and CEO of @coinbase, walked @davidsenra through the setup in a new conversation on his podcast:
"right now, I have probably seven or eight lead agents that are working on different areas that I'm focused on right now, and then one chief of staff, which kind of oversees them."
Each lead spins up its own helpers when it runs, "three to five different sub-agents or something." Seven or eight leads with three to five each is 21 to 40 sub-agents at full tilt, all reporting up to one person.
So he hit the wall you would expect: "I just couldn't keep all these walls of text."
The fix Coinbase is prototyping is an inbox of decisions, ranked by what moves the agents forward most. A product review with a couple of screenshots, a proposed next step, and he answers yes, no, or a few words by voice dictation, which he says is "faster than typing."
He laid out the structure on X on November 7: one lead agent per area of focus, running all the time, a chief of staff agent above them, and every change still going through the same human review and security checks.
That last part is why the inbox matters so much. If every change still needs a human yes, the whole system moves at the speed of that yes. Adding agents without fixing the queue just makes the pile of text taller.
His own verdict on the inbox: "that's reduced a lot of the cognitive overhead."
It also changes what a good manager of agents looks like. Less time writing prompts, more time making fast calls on a well-ranked list, which is a skill most executives already built in their email. The companies that copy his agent count without copying his queue will end up where he started, buried in text.
From @davidsenra's podcast with @brian_armstrong.
5
4
1,189
We book meetings for B2B companies on a pay-per-meeting basis.
You only pay for qualified prospects who actually show up.
If you're interested, book a call with us here:
introslot.com/revenueflow/ca…
161
Brian Armstrong says @coinbase is cutting its list of outside software vendors.
"October was just such a bad month for cyber incidents," he told @tbpn. "And don't worry, November will be even worse."
His fix: "some of it you can recreate in house and it just gives you more control."
Cheap code turns every vendor with access to your systems into a security question first and a feature second.
From @tbpn with @brian_armstrong.
1
765
We book meetings for B2B companies on a pay-per-meeting basis.
You only pay for qualified prospects who actually show up.
If you're interested, book a call with us here:
introslot.com/revenueflow/ca…
124
