What this covers

Patrick talks with Dylan Patel of SemiAnalysis about the supply and demand of tokens, starting from his own firm's numbers. AI spend went from tens of thousands of dollars last year to seven million this year. Examples include an energy analyst spending about $6,000 a day on Claude to map US power plants and transmission, and one person building a chip-imaging application that once took a team at a company like Intel. Mythos anchors the capability discussion. Patel reads it as proof that scaling laws still hold, and says its benchmarks imply a jump from roughly an L4 to an L6 software engineer in about two months.

The range is wide. On supply, he argues that memory makers' margins rise until prices clear the market, that CPUs are sold out because reinforcement-learning environments run on them, and that GPU useful life may stretch to seven or eight years. He also argues that the cost of a fixed capability level can fall a hundredfold while total spend still rises, because buyers move to the newer model. He puts Anthropic's gross margin at no lower than 72 percent, though that estimate is contested. He calls tokenomics the hardest area to model, since usage keeps outrunning forecasts.

The more speculative turns come later. His economist's grading of Bureau of Labor Statistics tasks leads to a "phantom GDP" argument, in which measured output understates value. He calls a software-only singularity a temporary blip because most of the world is physical. He expects few-shot breakthroughs in robotics within six to 18 months. And he predicts labs will sell their best models to ever fewer customers, partly to limit distillation, concentrating value among wealthy firms. He illustrates this with a hypothetical in which Ken Griffin commits $10 billion a year to tokens. He also claims that those who fail to capture value from tokens risk a permanent underclass.

Sharpest takeaway

Patel argues token demand is effectively unbounded because frontier models make implementation cheap and high-value, so GPUs, memory, fab capacity, and CPUs will stay sold out, and public backlash against AI is the main near-term risk to the buildout.

  • Frontier models unlock new economically valuable use cases, so demand is driven by capability rather than falling per-token cost
  • Supply chains react slowly: memory, fab capacity, and components cannot expand fast enough, so prices and margins rise
  • Concentrated access to the best models and a public backlash against AI are the main risks to how the buildout unfolds

The claims · ranked64 claims · weighted by value

This asset isn't compiled yet

You're seeing its claims, ranked. Compile it to build the argument threads, weight them, and check each claim against your library — the full view.

0.75

The cost to reach a fixed capability tier has fallen by one one-hundredth to one one-thousandth over time, with DeepSeek offering GPT-4-level capability at about one six-hundredth the cost, and the same quality could cost about a hundred times less a year from now.

factualhigh valueestablishednovelty 2/4durability 3/4· Dylan Patel

“The cost to hit a certain capability tier used to cost X. And now it costs one one-hundredth or one one-thousandth of that. DeepSeek, for example, on GPT-4 was one six-hundredth the cost.”

0.75

Information services sell for less than the value buyers extract from them, because a buyer only pays a dollar for information when it helps them make more than a dollar, so the buyer ends up earning more from the vendor than the vendor earns itself.

causalhigh valueestablishednovelty 2/4durability 3/4· Dylan Patel

“if I sell you information for a dollar, you're only buying it for a dollar because you know that information helps you make a decision that lets you make more than $1.”

0.75

In a capitalist economy, capacity can only be pulled from other uses through demand destruction via higher prices, so supply constraints are not rationed administratively and margins for memory makers rise until prices clear the market.

causalhigh valueestablishednovelty 2/4durability 3/4· Dylan Patel

“And the only way to steal capacity from somewhere else in a capitalist economy is demand destruction via higher pricing. We're not rationing stuff here.”

0.74

GPU useful life is extending beyond five years, possibly seven or eight, as three- and four-year-old Hopper and A100 clusters re-sign for further years at higher prices, so cloud gross margins are expanding.

factualhigh valuecontestednovelty 4/4durability 2/4· Dylan Patel

“There are clusters now re-signing. Three or four-year-old Hopper clusters re-signing for three or four more years.”

0.73

A software-only AI singularity will be a temporary blip rather than transformative for everything, because most of the world is physical and hardware, robotics, and actuators are much harder than software to improve.

forecasthigh valuecontestednovelty 3/4durability 3/4· Dylan Patel

“software-only singularity is just a blip and not like a, we do get everything else.”

0.73

The value created by tokens becomes better information sold at a discount, and it diffuses through the economy in ways that GDP statistics cannot capture, so official GDP understates the economic value being created.

causalhigh valuecontestednovelty 3/4durability 3/4· Dylan Patel

“Because it's not really something that you can capture in any GDP statistic.”

0.73

Users want the frontier model because it enables new economically valuable work, so falling costs of older GPT-4-class models do not drive overall demand, which is driven by new use cases.

causalhigh valuecontestednovelty 3/4durability 3/4· Dylan Patel

“Of course, no one gives a crap about GPT-4 class models. They want the frontier because the frontier lets them create the economically valuable things.”

0.69

A model whose quality matches today's Opus 4.6 or 4.7 could cost roughly one-hundredth as much a year from now, but total spend will still rise because users will buy the better, newer model instead.

forecasthigh valuecontestednovelty 3/4durability 2/4· Dylan Patel

“a year from now, my spend for the same exact quality of the model would probably be like 70K. I bet you it'll be a hundred times cheaper. Irrelevant because I'm going to be using a way, way, way better model”

0.68

Labs get compute-efficiency wins that cut the cost of a fixed capability level every few months, and scaling up then produces large capability jumps, so efficiency and scale compound.

causalhigh valuecontestednovelty 2/4durability 3/4· Dylan Patel

“Along the whole way, we're also getting these compute efficiency wins, which are as all this research compute that the labs are spending is actually turning into”

0.68

In reinforcement learning, the environment that scores model attempts runs on CPUs, while the model generating trajectories runs on GPUs or ASICs, and environments are growing from simple output-format checks to complex tasks such as editing CAD models in Siemens simulation software.

definitionhigh valueestablishednovelty 2/4durability 3/4· Dylan Patel

“those environments run on CPUs, they don't run on GPUs, they don't run on ASICs.”

0.64

Mythos is a materially larger model whose capability gain shows the scaling-law trend still holds, with more compute producing better models, while compute efficiency gains compound on top, so both bigger training and efficiency improve capability.

normativehigh valuecontestednovelty 2/4durability 3/4· Dylan Patel

“It's proof that the scaling laws still work. Everything about it shows the trend line continues”

0.63

With AI, implementation has become cheap and plentiful, so ideas, not execution, are the scarce input; only good ideas justify the token spend, and the valuable skill becomes choosing the right idea and selling what gets built.

normativehigh valuespeaker onlynovelty 3/4durability 4/4· Dylan Patel

“Now ideas are cheap and plentiful, but execution is very easy. So really only the good ideas are the ones that can justify the spend on super cheap implementation.”

0.61

Model release cadence has compressed to about two months because implementation of research ideas is now cheap and easy, so the bottleneck shifts to choosing which ideas are worth the token spend.

causalhigh valuespeaker onlynovelty 4/4durability 3/4· Dylan Patel

“Now ideas are cheap and plentiful, but execution is very easy. So really only the good ideas are the ones that can justify the spend on super cheap implementation.”

0.61

Anthropic aimed to reach an L4 software engineer level by end of 2025 and achieved it with Opus 4.6, while Mythos benchmarks suggest about L6 engineer ability, a jump in roughly two months.

factualhigh valuecontestednovelty 3/4durability 1/4· Dylan Patel

“if you look at Mythos, if you compare benchmarks, it's like an L6 engineer.”

0.57

Assuming Anthropic kept R&D compute flat and put all incremental compute toward inference, its gross margin is at least 72 percent, and likely higher once R&D compute is considered, with a leaked funding document earlier in the year showing 30-something percent margins.

factualhigh valuecontestednovelty 3/4durability 1/4· Dylan Patel

“their margins are at a floor of 72%. In reality, some of that incremental compute they've got probably went to research and development”

0.56

People who do not use more tokens, generate economic value from them, and capture that value will never escape a permanent underclass, as models grow in capability and resources concentrate.

normativehigh valuefringenovelty 2/4durability 2/4· Dylan Patel

“If you don't use more tokens, you'll never escape the permanent underclass.”

0.55

A SemiAnalysis economist used AI to grade about 2,000 Bureau of Labor Statistics tasks, found about 3% doable by AI now, and built a metric of AI-doable task cost to estimate deflationary effects, which he calls phantom GDP, since output can rise while measured GDP shrinks as costs fall.

normativehigh valuespeaker onlynovelty 4/4durability 2/4· Dylan Patel

“About 3% are doable now with AI. And so he's created this metric”

0.53

Labs will release their most capable models to fewer and fewer customers over time, partly to prevent distillation and partly because compute is scarce and the customers who pay most can build the most value.

forecasthigh valuespeaker onlynovelty 3/4durability 2/4· Dylan Patel

“And then you don't want people to distill your model so you don't release them broadly. You release them to fewer and fewer set of customers.”

0.53

Tokenomics is the hardest area to model because infrastructure costs and lab margins are well measured, but usage and adoption have repeatedly outrun forecasts, with Anthropic beating expectations month after month.

factualhigh valuespeaker onlynovelty 3/4durability 2/4· Dylan Patel

“But the usage and adoption is what's really difficult to model.”

0.53

To reduce fear, the AI industry should showcase present-day beneficial uses of AI and stop emphasizing future capability changes, because the public has no personal connection to lab employees and sees them as a cabal.

normativehigh valuespeaker onlynovelty 3/4durability 2/4· Dylan Patel

“they need to stop talking about how the capabilities are going to change the whole world constantly because then people are going to get fear of that capability because they have no connection.”

0.53

A single person at SemiAnalysis used a few thousand dollars of Claude tokens to build a GPU-accelerated chip-imaging application that labels materials across a chip, a task that previously required an entire team at a company like Intel.

normativehigh valuespeaker onlynovelty 3/4durability 2/4· Dylan Patel

“A few thousand dollars of Claude. The person previously worked at Intel and he said that was an entire team's job to build that and maintain that.”

0.53

A SemiAnalysis data-center energy lead spent about $6,000 a day on Claude and built a US power plant and transmission map with micro-region surplus and deficit analysis in weeks, which energy traders judged better than an incumbent with about 100 people working on it for a decade.

normativehigh valuespeaker onlynovelty 3/4durability 2/4· Dylan Patel

“XYZ company has a hundred people that have been working on this for a decade.”

0.53

Vision-language-action robot models are inefficient in data and cannot scale their data fast enough, so they are unlikely to be the approach that ultimately scales, and few-shot learning breakthroughs in robotics could come within six to 18 months.

forecasthigh valuespeaker onlynovelty 3/4durability 2/4· Dylan Patel

“VLAs, Vision Language Action Models, which is very popular right now, is probably not going to be the thing that ultimately scales beyond.”

0.53

As frontier model access narrows, token usage and the value it creates will concentrate among fewer companies, with wealthy firms able to secure early access and crush competitors, as in a hypothetical where Ken Griffin commits $10 billion a year to tokens.

forecasthigh valuespeaker onlynovelty 3/4durability 2/4· Dylan Patel

“So then the question becomes, where does this cycle end where token usage and therefore the benefits of those tokens”

0.53

CPUs are completely sold out because reinforcement learning environments run on CPUs and because AI-generated apps are deployed on cloud CPU instances, so CPU demand is skyrocketing alongside GPU demand.

factualhigh valuespeaker onlynovelty 3/4durability 2/4· Dylan Patel

“CPU-wise, all these reinforcement learning environments, plus all the slop code you and I are generating”

0.52

AI commoditizes information services like it commoditizes software, so firms that do not keep moving up the quality bar will lose their edge, while those that move fast and keep customers will grow faster.

causalhigh valuespeaker onlynovelty 2/4durability 3/4· Dylan Patel

“If I don't move up the bar, then I will be commoditized. If I don't move fast enough, I will also lose my edge.”

0.51

Tier-1 labs can charge higher gross margins, around 70 percent or more, while tier-2 labs will sell out at lower margins around 50 percent, so lab margins keep expanding until hardware suppliers raise prices.

forecasthigh valuespeaker onlynovelty 3/4durability 2/4· Dylan Patel

“Anthropic may get to charge 70 plus percent gross margins. But if OpenAI hits it next, they charge 50% gross margins.”

0.51

Margins are high at NVIDIA, around 75 percent gross, and rising in memory, while chip companies pay large prepayments to suppliers, which raises return on invested capital even where gross margins are flat.

factualhigh valuespeaker onlynovelty 3/4durability 2/4· Dylan Patel

“NVIDIA still charging 75 or whatever percent gross margin.”

0.51

Niche supply chains such as copper foil, glass fiber for PCBs, and lasers are very tight, and upstream semiconductor wafer fabrication equipment has risen a lot but is still underappreciated as a bottleneck.

factualhigh valuespeaker onlynovelty 3/4durability 2/4· Dylan Patel

“These are things that are well understood and niche supply chains, but they're very, very tight.”

0.50

By the end of the year the economy could spend about $100 billion on models at the Opus 4.6 tier, up from about $40 billion now, which Dylan calls a linear extrapolation driven by diffusion rather than model improvement.

forecasthigh valuespeaker onlynovelty 3/4durability 1/4· Dylan Patel

“by the end of the year, let's say a 4.6 Opus tier model, the economy would spend $100 billion on. I don't think that's unreasonable.”

0.50

Memory capacity grows only 20 to 30 percent a year, with true incremental DRAM supply not arriving until late 2027 or 2028, so DRAM prices could still double or triple from current levels.

forecasthigh valuespeaker onlynovelty 3/4durability 1/4· Dylan Patel

“DRAM will double or triple from here still, because that's how much capacity is required.”

0.49

When demand exceeds supply, what matters is having an enterprise contract and a lab representative who can raise rate limits, because tokens go to whoever can pay for them.

causalhigh valuespeaker onlynovelty 2/4durability 2/4· Dylan Patel

“What really matters is having an Anthropic rep and having an enterprise contract with them and getting the rate limit increases that you need.”

0.49

AI industry leaders such as Sam Altman and Dario Amodei should stop doing interviews that make people hate them, should show uplifting uses of AI now, and should stop emphasizing future capability changes that frighten people.

normativehigh valuespeaker onlynovelty 2/4durability 2/4· Dylan Patel

“they need to start showing uplifting things that can be done with AI. Three, they need to stop talking about how the capabilities are going to change the whole world constantly”

0.49

Tokens generate more economic value than they cost, so businesses that create little value per token, such as a weak SaaS startup, will eventually be priced out of token access.

normativehigh valuespeaker onlynovelty 2/4durability 2/4· Dylan Patel

“the shitty SaaS startup in SF who is using Claude to generate their software product is not necessarily actually creating a ton of value and therefore they're going to get priced out of tokens soon enough.”

0.48

Anthropic's Mythos may be the biggest step up in model capability in two years, and although it costs more per token it uses far fewer tokens per task, making it cheaper than Opus 4.6 for most tasks.

factualhigh valuespeaker onlynovelty 3/4durability 2/4· Dylan Patel

“Anthropic Mythos is more expensive as a model, but it spends a lot less tokens to do the thing. And therefore, it is actually cheaper in most tasks than 4.6 Opus”

0.48

Anthropic chose not to broadly release Mythos over concern about its impact, released a worse Opus 4.7 instead, and the model card states the company preferentially made it worse at cyber capabilities.

factualhigh valuespeaker onlynovelty 3/4durability 2/4· Dylan Patel

“And they explicitly said in the model card, hey, we actually preferentially made it worse at cyber.”

0.48

Anthropic could double Opus pricing and most users would keep paying, but that would not solve its capacity shortfall, which shows that price is not the constraint on demand.

factualhigh valuespeaker onlynovelty 3/4durability 2/4· Dylan Patel

“They could double their pricing on Opus and I would continue to pay and most users would continue to pay. I bet that wouldn't solve their humongous capacity problem that they have.”

0.48

Next-generation AI racks contain about 120 FPGAs each, making FPGA names a notable part of the AI infrastructure supply chain.

factualhigh valuespeaker onlynovelty 3/4durability 2/4· Dylan Patel

“there's 120 FPGAs per next-generation rack, AI rack.”

0.47

At the current rate SemiAnalysis spends about $7M a year on Claude Code versus roughly $25M in salary expense, over 25% of salary, and if the trajectory continues spend would exceed 100% of salary by year end.

forecasthigh valuespeaker onlynovelty 3/4durability 1/4· Dylan Patel

“we're spending $7 million a year now on Claude Code at the current rate, versus our salary expense being in the neighborhood of $25 million.”

0.46

Mythos is potentially the largest step up in model capability in two years, and Anthropic withheld it for broad release while offering a deliberately weaker Opus 4.7 with reduced cyber capability.

factualhigh valuespeaker onlynovelty 2/4durability 2/4· Dylan Patel

“Mythos is potentially the biggest step up in model capabilities in two years.”

0.45

SemiAnalysis's AI spend went from tens of thousands of dollars last year to a roughly $7 million run rate this year, after spend inflected in January following Opus adoption and an enterprise Anthropic contract.

factualhigh valuespeaker onlynovelty 2/4durability 1/4· Dylan Patel

“This year, the spend has just skyrocketed. And it really started in late December with Opus.”

0.45

Dylan expects large-scale protests against Anthropic and OpenAI within three months, because AI is less popular than ICE and politicians, and downstream business changes will spread fear.

forecasthigh valuespeaker onlynovelty 2/4durability 1/4· Dylan Patel

“I think there will be a large-scale protest against Anthropic and OpenAI. People hate AI. AI is less popular than ICE, less popular than politicians.”

0.45

Over the next six to 18 months, Patel expects breakthroughs in robotics that enable few-shot learning from a pre-trained robot model, making robots useful for niche tasks such as chalkboard cleaning rentals, which will be deflationary and keep token demand growing.

forecasthigh valuespeaker onlynovelty 2/4durability 1/4· Dylan Patel

“in the next six to 18 months, we'll start seeing real breakthroughs in robotics that enable few-shot learning”

0.45

TSMC capex is about $56 billion this year and may reach around $100 billion by 2027 or 2028, which would ripple into equipment suppliers such as Lam Research, Applied Materials, ASML, and MKSI, and TSMC is raising prices only in single digits.

forecasthigh valuespeaker onlynovelty 3/4durability 1/4· Dylan Patel

“TSMC CapEx this year, they say $56 billion.”

0.44

Supply chains are more complex than ever and lead times are longer, but industries have dealt with 18-month lead times before, and building incremental supply did not used to take years.

causalestablishednovelty 2/4durability 2/4· Dylan Patel

“our supply chains now are more complex than ever. And the things we're building are more complex than ever. And therefore the lead times are longer.”

0.44

Cloud-layer and hardware-layer margins are rising, with NVIDIA still charging about 75% gross margin, because GPU prices and renewal rates are increasing even as capacity is tight.

factualhigh valuespeaker onlynovelty 2/4durability 2/4· Dylan Patel

“So margins are expanding in the cloud layer. Margins are extremely healthy on the hardware layer with NVIDIA still charging 75 or whatever percent gross margin.”

0.43

Anthropic's annualized revenue grew from about $9 billion to $35-40 billion, with perhaps $40-45 billion by air date, while its compute grew much more slowly.

factualhigh valuespeaker onlynovelty 2/4durability 1/4· Dylan Patel

“Anthropic has gone from $9 billion revenue to what they're at $35, $40 billion.”

0.31

Robotics will be an explosion in physical goods and deflationary effects, and that will sustain token demand growth, which Dylan does not expect to slow.

forecastspeaker onlynovelty 2/4durability 2/4· Dylan Patel

“There is going to be some way to large-scale pre-trained robot models”

0.31

The average person does not know anyone at Anthropic or OpenAI and views these labs as a small cabal of about 5,000 people trying to automate jobs and destroy society, which makes fear of AI capability grow.

factualspeaker onlynovelty 2/4durability 2/4· Dylan Patel

“they just view them as a sneaky cabal of 5,000 people at this company that are going to change the world and automate all the jobs and destroy society.”

0.31

Non-technical people using Claude Code spend thousands of dollars a day at times, and individual user spend oscillates widely, though firm-wide spend keeps rising.

factualspeaker onlynovelty 2/4durability 2/4· Dylan Patel

“People who have never coded before are using Claude Code and spending thousands of dollars, sometimes a day. And it oscillates.”

0.31

People will increasingly blame AI for deep-seated global problems, and politicians or influencers will weaponize AI against people, so public fear will grow as Anthropic revenue scales and causes downstream business changes.

causalspeaker onlynovelty 2/4durability 2/4· Dylan Patel

“They'll start blaming more and more of their own problems and things that are global, have been deep-seated problems for a long time.”

0.27

Anthropic's compute limits mean it cannot grow as fast as demand, and Dario Amodei had previously criticized OpenAI for being too aggressive on compute, a position Anthropic now seems to regret.

factualspeaker onlynovelty 2/4durability 1/4· Dylan Patel

“Dario used to gloat about how OpenAI was being too aggressive on compute and Anthropic was more sensible”

0.27

In the near term, businesses are essentially arbitraging tokens by choosing which direction to point them, but in three or four years models themselves will know how to direct tokens for maximum value.

forecastspeaker onlynovelty 2/4durability 1/4· Dylan Patel

“the business is actually just arbitraging tokens. The tokens are amazing, but let's figure out what direction to point them in. And then three or four years from now, the model will know what to do with the tokens”

0.27

Dylan and a colleague begged an Anthropic co-founder for Mythos access, and the co-founder pretended the model did not exist even though Anthropic had already priced it for selected users.

factualspeaker onlynovelty 2/4durability 1/4· Dylan Patel

“myself and a buddy of mine, Leopold, being on our knees in front of an Anthropic co-founder, begging him for access to Mythos and them pretending it doesn't exist.”

0.26

Sophisticated investment funds with their own data teams will still purchase data from specialist vendors, because buying and building on top is cheaper than building from scratch and because speed and nimbleness are an advantage.

causalspeaker onlynovelty 2/4durability 2/4· Dylan Patel

“And more likely they'll just buy the data from us. And it's cheaper for them to buy the data from us and then build on top of it”

0.26

Compute for frontier training is largely set by commitments already signed and changes only slowly, while top researchers cost tens of millions of dollars, so implementation has become the key lever that is now cheap.

causalspeaker onlynovelty 2/4durability 2/4· Dylan Patel

“It's like set in stone what you've already signed and there'll be delays and shifts”

0.26

If a rich buyer like Ken Griffin of Citadel committed billions in advance to first access to frontier tokens, he could outcompete rivals in markets, and similar advantages in cyber or information services could compound.

forecastspeaker onlynovelty 2/4durability 2/4· Dylan Patel

“He goes and signs a deal with Open Air Anthropic that's like, yeah, I'm going to get access to your models and I'll buy the first 10 billion dollars worth of tokens each year.”

0.26

TSMC is raising prices only in single digits while memory makers have raised them by triple digits, so the question is whether TSMC will capture the full value of its position in a tight market.

factualspeaker onlynovelty 2/4durability 2/4· Dylan Patel

“they're not raising prices fast because they're good people, it seems like. Single-digit price increases instead of triple-digit price increases like the memory guys have had.”

0.26

Rising fab capex compounds through the supply chain like a tailwhip, so each increase in TSMC's spending makes shortages in downstream equipment suppliers grow harder and faster.

causalspeaker onlynovelty 2/4durability 2/4· Dylan Patel

“The tailwhip, it just gets whipped harder and harder and harder.”

0.23

Much of the spend increase came from non-technical staff, such as the firm's president Doug O'Laughlin, using Claude Code for coding, which slowly converted the whole firm to heavy AI use.

causalspeaker onlynovelty 1/4durability 1/4· Dylan Patel

“That included Doug O'Laughlin, who's president, and he's very much leading the charge in the sense of non-technical people using AI for coding. And so he's basically pilled the whole firm slowly over time.”

0.22

Anthropic made a large training jump with Mythos, while OpenAI is taking a more principled approach of smaller scaling steps, and release cadence across labs will keep accelerating through the year.

factualspeaker onlynovelty 2/4durability 1/4· Dylan Patel

“Anthropic really went for a huge jump. We'll see better and better models throughout the year.”

0.21

Heavy users with capital should buy enterprise pay-per-token Anthropic contracts rather than consumer subscriptions, since subscriptions face rate limits, and then use the tokens on the highest-value tasks.

normativespeaker onlynovelty 1/4durability 2/4· Dylan Patel

“you should get a freaking enterprise Anthropic subscription where you pay per token, not with these subscriptions, because then you won't get rate limited much.”

0.20

Sam Altman had a Molotov cocktail thrown at his house twice in two weeks, and online news comments cheered it on, which Dylan cites as evidence of growing hostility toward AI.

factualspeaker onlynovelty 1/4durability 1/4· Dylan Patel

“Sam Altman had a Molotov cocktail thrown at his house twice in two weeks. People were cheering it on.”

0.20

Patrick reported that once Opus 4.7 released he immediately wanted to use it and could not stop wanting the newest model, even though 4.6 had satisfied him for weeks, illustrating users' insistence on the frontier.

factualspeaker onlynovelty 1/4durability 1/4· Patrick

“I couldn't think about using 4.6 anymore now that 4.7 is out. I was perfectly happy with 4.6 for the last many weeks.”