If Your Tokens Aren't Worth 4X Inside Your Product, What Are You Even Doing?

If Your Tokens Aren’t Worth 4X Inside Your Product, What Are You Even Doing?

The sharpest argument about AI business models this week came in two sentences. Vanta’s CEO Christina Cacioppo, quoted by Patrick O’Shaughnessy, said her company eats its token costs: “If you try to charge people based on tokens, they are very quick to say, ‘you’ve just marked up a Claude token 4X. Are you 4X as good as Claude?’” You put yourself in competition with the lab, she argues, and that is a fight you lose. Parker Conrad disagreed, hard. His reply: you need to prove — to yourself more than to investors — that intelligence is more valuable inside your product than on top of it. If your tokens aren’t worth four times inside your product, you are not building a product at all. You are just reselling the labs.

This is the debate that decides which AI companies survive the next pricing war. Cacioppo’s move — absorb the token cost, charge for the outcome — is a distribution play: remove the line item customers argue about and compete on workflow. It works when your software already commands a price and tokens are a rounding error. Conrad’s counter is a product play: the right question is not how do we hide the token cost but why is our token 4X better than a raw lab token. His answer has to be the product itself — the context it carries, the workflows it automates, the data flywheel it spins. If none of that multiplies the intelligence, you are, in his framing, a middleman with a markup, and middlemen get compressed.

Read it as a test for any AI startup, including the ones in today’s pile: Beacon (below) is valuable precisely because it makes your sessions compound inside your agents — that is intelligence worth more inside the product. WisprFlow’s content engine (also below) passes the test on distribution, not tokens: the moat is the founder’s trust, which no lab can undercut. And watch how Cacioppo’s position quietly concedes Conrad’s premise: Vanta eats the tokens because Vanta’s value was never the tokens. The fight is over where the 4X lives. Put it anywhere except inside your product and you lose it.

• Track gross margin on token spend at full utilization as the core unit-economics metric for any AI feature — not blended margin, margin when the heaviest users are hammering it.

• Never let the customer do the arithmetic “you marked up a lab token 4X” unless you can point at exactly what the 4X buys: proprietary context, workflow automation, a data flywheel.

• Cacioppo’s alternative is legitimate: if tokens are a rounding error next to your seat price, eat them and compete on outcomes. The worst position is the middle — passing costs through and adding no intelligence.

Diffusion Finally Grows Up: Mercury 3 at 1,068 Tokens a Second

Inception dropped Mercury 3 on October 19th, and it is the first diffusion language model that can credibly claim to sit at the frontier’s speed tier on quality too. The company’s numbers: more than 40 times the size of Mercury 1, generating at 1,068 tokens per second on NVIDIA GPUs, beating GPT-6 Luna and Claude Haiku 5.5 on 8 of 12 benchmarks while running four to eight times faster. Context window jumps to one million tokens. Launch pricing is half the list rate — ten cents per million input tokens, thirty-five cents per million output — and Deedy notes OpenRouter shows it effectively two to four times faster and cheaper than Luna/Haiku in real use.

Why it matters: for a year, diffusion models were the cheap, fast sidekick — great for latency, suspect on quality. Mercury 3 is Inception selling one as a competitor, and the scaling claim is the interesting part: the company says the same scaling laws that drove autoregressive progress hold for diffusion, with throughput essentially unchanged as the model grew 40X. If that holds up under independent testing, the frontier splits into two architectures and every latency-sensitive workload — browser agents, voice, real-time search — gets a new default. The caveats are real: these are vendor benchmarks, no outside group has published results yet, and the widest lead came in coding (51.1 on SWE-Bench Atlas vs 33.3 for Luna) while Haiku 5.5 still won DeepSWE. TinyFish AI and Octen AI’s early production numbers (half of 50 live navigation tasks completed; 0.9s median generation) were also shared by Inception, not independently verified. Still: diffusion scaling without losing speed was the open question, and this is the first serious answer.

• Architecture is back as a differentiator: parallelism beats sequence when latency is the product.

• Watch the migration deadline — Mercury 2/2.5 close to new users, existing users keep access until November 15, 2026. Test Mercury 3 on your own workloads before then.

• Treat vendor benchmarks as a hypothesis, not a result: the coding lead is promising, the missing independent evals are the gap.

Beacon: Your Coding Agents Finally Remember You

Every coding session teaches your agent something — how you debug, how you structure a fix, what “done” looks like in your codebase — and then the session ends and all of it evaporates. You re-explain everything next time, in every tool. Justin D’Souza’s bet, backed by a $5.5M seed round, is that the next leap in AI coding won’t come from OpenAI or Anthropic at all. It will come from a new feedback loop that learns from you. Beacon is open-source, self-improving memory for coding agents: it continuously distills your sessions into reusable skills that work across Claude Code, Cursor, Codex, and 20+ harnesses.

The framing is the product. “Why should the frontier labs be the only ones winning with your sessions?” Every prompt you write trains their models; Beacon keeps the learning on your side of the fence, as portable skills rather than locked-in model weights. That is the same “inside” logic as the lead story: the durable asset is not the rented intelligence, it is what accumulates inside your loop. If it works, the moat for a coding agent stops being which lab powers it and starts being how many of your sessions it has absorbed. Open-source and free, at github.com/Asymptote-Labs/agent-beacon per the launch post. The risk is the usual one for memory systems: distilled skills are only as good as the sessions they learn from, and garbage debugging habits will compound just as faithfully as good ones.

The Founder-Led Content Engine, Deconstructed

WisprFlow’s marketing hire documented exactly how the company builds its founder-led content engine, and Jamie Quint’s summary is the rare growth thread with numbers attached. The thesis is blunt: brand is the moat. Products and features are easy to copy; trust and audience relationships are not. Paid and organic run together — paid rapidly tests messaging and drives conversions while organic builds trust, familiarity, and lowers acquisition cost. YouTube is the highest-leverage organic channel: videos compound views for years, long-form builds deeper trust, and a founder can simply document what they are already doing.

Two details deserve attention. First, the cost structure is published, which almost never happens: a head of content at ten to sixty thousand dollars a month, a videographer at seven to fifteen, a YouTube team at five to twenty-five, short-form editors at one to three thousand. Early-stage companies start with one generalist. Second, the operating principle: treat content like a product roadmap — plan around launches and milestones, capture the unexpected moments too — and remember that packaging beats production quality. Titles, thumbnails, pacing, and storytelling drive performance, not camera gear.

JakeOnGrowth’s reply is the necessary corrective: none of this is a secret. Hector is “smart and extremely hungry,” executing and testing ideas at a fast pace — and most people who watch the content will feel like they cracked it and then do nothing. The playbook is public; the pace is the moat. Read alongside Conrad’s lead: the founder’s trust is exactly the kind of value that lives inside the company, where no lab and no competitor can undercut it on price.

• Document, don’t perform: the founder’s actual work is the content; the team’s job is packaging.

• Run paid as a laboratory (fast message tests) and organic as the compounding asset (trust that lowers CAC over time).

• Onboard the content team into the founder’s narrative — background, struggles, motivations — before handing them a camera.

She Left Apple to End Chronic Suffering

Praneeta Pujari left Apple to build UniqueHuman: proactive healthcare agents that handle every step of care for people with chronic conditions. The mission statement is deliberately enormous — “end chronic suffering” — and the starting point is deliberately narrow: migraines. That wedge choice is the whole strategy. Chronic care is a coordination problem wrapped in a monitoring problem wrapped in an empathy problem; an agent that listens early, acts before the crisis, and shepherds a patient through every step is worth more than ten symptom-checker chatbots. Starting with migraines — recurrent, trackable, trigger-laden, badly served by episodic care — gives the agent a tight feedback loop before it expands to the rest of chronic disease. One to watch: healthcare is where “proactive agent” stops being a demo and starts being a duty of care.

The TiDB Cloud Breach: 10.6 TB and a Warning

Arnav Gupta’s two-word reaction — “Holy shit” — was doing a lot of work. The ransomware group FulcrumSec claims to have exfiltrated roughly 10.6 terabytes from TiDB Cloud, PingCAP’s managed database service, and has released a ~200 GB first sample. PingCAP has since disclosed the incident itself: an unauthorized party used stolen service-account credentials to access internal systems between September 9 and October 7, copying customer data and obtaining source code. Exported records reportedly include TiDB Cloud account details, SQL editor text, cluster identifiers, and business and billing information. PingCAP says it has blocked the access paths and that there is no evidence data stored inside TiDB databases was affected; customers like Bitget say they used the service only for public trade logs. The uncomfortable part for everyone else: the breach wasn’t a product vulnerability — it was credentials. The database company got breached the boring way, which is how nearly everyone gets breached.

Two Dispatches from the Agent Age

1. The good-night agent. Gabriel told Claude to work extra hard and ask all questions upfront so he could go to sleep while it kept working. After giving it his preferences, it wished him a good night. File under: the UX of delegation is becoming the product. The model that asks the right questions before you sleep is worth more than the model that writes the fastest code while you watch.

2. Jev, spotted in production. Thomas Christensen caught Slack suggesting a salad emoji for the word “lunch” in a status update. The smallest possible sighting of the ambient-AI era: the assistant no longer waits to be invoked, it just quietly finishes your sentence. Cute at the emoji level; load-bearing at the workflow level.

Your Ask Is Your Valuation

Sajith Pai watched a founder make the rookie mistake in real time — and the founder didn’t even know he’d made it. Asked how much he was raising, the founder quoted his number. Pai told him the valuation was high for the revenue. The founder protested: “I never said anything about valuation.” Pai’s answer: you didn’t have to. The moment you quote an ask, the VC multiplies it by four to five — early rounds run 20 to 25 percent dilution — and internalises the result as the valuation you expect. Your ask effectively suggests your valuation.

The deeper mechanic: valuation at seed is more art than arithmetic, but VCs anchor on one thing — what the next investor will pay. The rule of thumb Pai lays out: the next round’s pre-money should be at least twice this round’s post-money, ideally within 15 to 18 months. If the VC doesn’t believe the next investor will value you at ₹500 crores-plus, they won’t value you at ₹250 crores today. So the ask isn’t derived from what you need; it’s derived from the valuation you can defend at the next fundraise, which is derived from whether your sector has what he calls capital legibility — how receptive investors currently are to your category. The practical move: pick the raise size last, not first. Decide the valuation story you can credibly tell eighteen months from now, halve it, multiply by your dilution tolerance, and that is your ask.

• Never quote an ask without doing the dilution math yourself: ask ÷ 0.20–0.25 = the valuation you just proposed.

• Price backwards from the next round: can you clear a 2X valuation bump in 15–18 months? If not, the ask is too big or the story is too small.

• Sector timing is a valuation input, not just a pitch input — capital legibility moves the 2X bar.

Commit or Move On: The Stuck-Founder Tax

Rohit Mittal says his conversations with stuck founders under $5M ARR are “like a broken record.” Most don’t have their hearts in running the company anymore — but they spend months, sometimes years, dilly-dallying before deciding what’s next. His prescription is binary: commit to going all-in until you hit escape velocity or prove it isn’t working, and the moment you decide it isn’t working, move fast on the options. Because the opportunity cost of “exploring options” is the number nobody prices.

He speaks from the scar tissue: at Stilt, he wishes he had acted a year faster. “That one year changed everything for us. But once we decided, we moved as fast as we could. We knew selling for a slightly higher price wouldn’t change anyone’s life. Closing a deal and working on something new might.” The uncomfortable insight is aimed at founders who know the answer and won’t say it: the founders who moved to fast-growing AI companies are, by his observation, doing much better than the ones who knew it wasn’t working and stayed anyway. Make a decision. Commit to it. Close the deal even if it isn’t optimal. The long game rewards velocity of truth, not perfection of outcome.

Study More, Earn Less? The PLFS Inversion

“The longer you study, the lesser is your likelihood to get a job.” Aviral Bhatnagar’s post, built on India’s Periodic Labour Force Survey 2025 data, states the inversion plainly: education and employment now have an inverse relationship. The old reassurance — padhoge likhoge banoge nawab, study and you’ll become a king — reads, in his rendering, as its opposite. The chart behind the post shows unemployment rising with educational attainment: the more years of schooling, the worse the odds.

Read it carefully and it’s not an argument against education — it’s an argument about what the labour market currently values. Degrees signal queue position for a shrinking set of formal jobs while the economy’s growth concentrates in roles where credentials matter less than demonstrable skill. The graduate unemployment bulge is what happens when the education system optimises for the old exam and the market hires for the new one. Pair it with Hriday Kadam’s career advice below — “visibility beats effort… being legible is the multiplier” — and the through-line sharpens: in a market where the degree no longer clears, what’s legible about you is what gets you hired.

The Influencer Is Dead. The Artisan Era Is Here.

Dan Koe’s long essay opens with an observation that should unsettle anyone building a personal brand: the people clearing million-dollar launches now are non-technical mothers in their forties selling simple digital products — a thirty-day challenge, a waitlist, validated content, a launch to the waitlist and an Instagram account. Nobody relates to the influencer anymore — the out-of-touch luxury lifestyle, the unreachable body. People relate to someone in their shoes, one or two steps ahead. The influencer optimises for attention itself; the personal brand optimises for monetisation. Koe argues both are being displaced by a third species: the artisan.

The artisan uses social media to subsidise a body of work — a book, a philosophy, research, ideas that live on in the reader’s head. James Clear, Jordan Peterson, Alan Watts, Marcus Aurelius: the examples Koe reaches for are people whose work compounds long after the algorithm forgets them. The practical half of the essay is a system for becoming one: build an “idea museum” — a ruthless swipe file of notes and inspirations — because your perspective is the one asset AI can’t generate for you. (He demos it in his own tool, Eden, but the habit matters more than the app.) The line that lands hardest: in a world where anyone can generate fifteen hundred nice-sounding words in ten seconds, a perspective only you could hold is the last moat. Read it next to this edition’s through-line: the influencer rents attention; the artisan owns the work. Inside, again, is where the value is.

• Stop optimising for the platform’s incentives; optimise for the body of work the platform subsidises.

• Your edge is an opinion nobody else’s life could have produced — but it’s invisible until you build a museum for it. Capture ideas daily or they evaporate.

• “You are the niche”: people learn best from someone one or two steps ahead of them, not from the unreachable expert.

Hard Work Is Table Stakes. Being Legible Is the Multiplier.

Hriday Kadam’s five-line career memo is the most quotable thing in today’s pile, and it pairs uncomfortably well with the PLFS inversion above. The notes: visibility beats effort — work nobody can see counts as zero at review time. Your writing is your reputation — every doc and message is a sample of your thinking. Seniors aren’t faster typists; they’ve just seen this bug before. The legacy codebase is never as bad as the blog post said — the post was written by someone who left. And shipping something small every week beats the giant launch that slips three quarters. The kicker: hard work is table stakes; being legible is the multiplier. In a labour market where the credential no longer clears, legibility — being readable to the people who decide — is the whole game.

You Are the Dataset

Ashwini Asokan’s reaction to Erik Newsham’s piece on AI data brokering is the rare tech-insider post that reads as genuinely alarmed: “mindblowing and mindblowingly scary.” The piece’s premise: anyone, anywhere, can capture a photo, a video, a fragment of someone’s life without them knowing — and package it and sell it to the labs for training. “Forget data breaches the way we have known them from our systems. We are data. Our physical lives are data. That someone is capturing without you knowing and selling it to labs.”

The unsettling mechanics underneath: most labs, as she notes quoting Ilya, have no idea what data they’re actually getting from the brokers. The supply chain runs from street-level capture to broker to training run with almost no provenance, and the money in it, per Newsham, is “absurd.” The line that should sit with you: the onus of protection has moved to the individual. Data privacy used to mean securing your systems; now it means your physical existence is the asset and nobody asked your permission. Pair it with the TiDB breach above and the picture sharpens — the industry is leaking from both ends, the cloud account and the camera on the street.

Brief: Add a Sense of Urgency

Ferriss’s latest video — “Life is Short — How to Add a Sense of Urgency” (9:33) — is nine and a half minutes on manufacturing the one resource motivation can’t buy: a deadline that feels real. File it next to Rohit Mittal’s “commit or move on” above. The founders who dilly-dally and the careers that drift share a root cause — no felt urgency — and Ferriss’s whole project has always been engineering it artificially before life engineers it for you.

The Unread Pile — Edition No. 27 — built from 16 new saves by @RAKSHITLODHA, Oct 20–21, 2026. The value has to live inside. See you tomorrow.