Solo founder of Trawly. Father of two (3 and 6), in Sydney. Builds mostly 9pm to midnight. Tracks AI and Claude Code, building and GTM, deal flow, biotech, and raising thinkers. The handwritten notes are his.
Trawly
Friday 24 July 2026 · Daily Trawl №164
Morning David, trawled up 571 posts for you. 14 keepers. Here's the catch:
🐋 The Big One
Anthropic halved Batch API pricing overnight and added scheduled runs. Overnight AI pipelines just got 50% cheaper
No keynote, just a changelog entry and a post at 1am AEST: batch jobs now run at 75% off standard rates (up from 50%), every model tier including Opus is eligible, and a new scheduled-runs option lets you queue a batch tonight with guaranteed completion by a deadline you pick. Third-party cost dashboards were showing the new rates within the hour.
C
Claude
@claudeai
Batch API update, live now:
→ 75% off standard rates (up from 50%)
→ all model tiers, including Opus
→ scheduled runs: queue tonight, guaranteed done by the deadline you set
No code changes. Existing batch jobs get the new pricing automatically.
❤️ 29,304 · 🔖 8,112
View on X →
🎯 Why it mattersTrawly's whole pipeline is an overnight batch, and your cost sheet is built on the old rates. At half the price, the Opus-everywhere compile you parked on cost grounds is back on the table, and the completion guarantee removes the 6am delivery risk that kept you off batch in the first place. Worth re-running the numbers before the week is out.
+Save · ↗Share
🧠 AI Frontier
AI Engineering
An engineer reverse-engineered Claude Code's API calls: 20,000 tokens of overhead in every request
Sam Lazlo logged 1,200 of his own sessions over a week. Every call carries roughly 20K tokens of system prompt, tool descriptions, and few-shot examples before your input reaches the model. Anthropic acknowledged it on Discord within four hours; a 'lite mode' ships next month. The tools.json config in the post strips about 6K immediately.
Sam Lazlo
@samlazlo
I logged 1,200 of my own Claude Code sessions over a week.
Every. Single. Call. carries ~20,000 tokens of overhead before your code touches the API.
System prompt: 7K. Tool descriptions: 9K. Few-shot examples: 4K.
Full breakdown + the tools.json config that strips 6K immediately ↓
❤️ 2,412 · 🔖 3,187
View on X →
🎯 Why it mattersYou build with Claude Code daily, so 20K of overhead per call adds up fast. The tools.json config in the post is a copy-it-in-this-morning fix. The wider signal: the layer you build on still has fat in it, and it took an outsider to surface it.
+Save · ↗Share
Local Models
Berkeley open-sourced Apex-30B: 60 tokens a second on a Mac mini, near-frontier coding scores
The new 'sparse-recurrent attention' technique drops VRAM by 4x without quality loss. It hits 60 tok/sec on a 16GB M4 and lands within a few points of frontier models on HumanEval+. A Hugging Face researcher is pushing back on the agentic-task numbers; the M3 memory caveat buried in the post is the bigger one to read.
bair.berkeley.edu · 2d ago
Apex-30B: Sparse-Recurrent Attention Brings Frontier Coding to Consumer Silicon
Apex-30B: Sparse-Recurrent Attention Brings Frontier Coding to Consumer Silicon. The BAIR Blog. We release Apex-30B, a 30B-parameter model that…
Read the article →
🎯 Why it mattersTrawly runs on cloud APIs today; local inference is the curve every solo AI founder watches. Apex-30B isn't ready to carry a whole pipeline, but a single low-stakes step could move local on hardware you'd already own. Validate the agentic-task softness in the rebuttal thread first.
+Save · ↗Share
Agent Skills
Garry Tan released 'GStack', YC's open-source library of 47 agent skills, and it's exploding on X
47 prompts and workflows for solo-founder tasks, each a Markdown file with prompt, examples, and an evaluation rubric. Skills include a customer-interview synthesizer, cold-email tester, fundraise-deck critic, and GTM-brief generator. The pitch: fork it, modify it, contribute back.
Garry Tan
@garrytan
Today YC is open-sourcing GStack: 47 agent skills every solo founder should run.
→ Customer interview synthesizer
→ Cold email tester
→ Fundraise deck critic
→ GTM brief generator
→ Hire spec writer
…42 more.
Fork it, modify it, contribute back. github.com/ycombinator/gstack
❤️ 18,229 · 🔖 14,466
View on X →
🎯 Why it mattersThe closest thing to a 'solo founder agent kit' you'll see this year. Three of the 47 hit your week directly: customer-interview synthesizer, cold-email tester, GTM-brief generator. The skill-as-Markdown format itself is a reusable pattern for your internal automations.
+Save · ↗Share
🔨 Builder Life
Solo Workflows
Karpathy open-sourced 'eureka', his personal solo-founder workflow
A 200-line Python orchestrator he runs almost everything through, from research to product specs. Claude for heavy lifts, Haiku for triage, and a small feedback loop that self-improves over time. The walkthrough video drops at noon PT. Short enough to read in 15 minutes; the architecture is the part to study.
Andrej Karpathy
@karpathy
Open-sourced eureka today, the small workflow I run almost everything through.
200 lines of Python. Claude for heavy lifts, Haiku for triage, a tiny feedback loop that self-improves over time.
Repo + 20-min walkthrough video at noon PT. Steal it, fork it, make it yours. github.com/karpathy/eureka
❤️ 38,204 · 🔖 12,841
View on X →
🎯 Why it mattersMost agentic products converge on a critique-then-revise shape, and Karpathy's is the cleanest version of it you'll read this year. The orchestrator pattern at the top of the repo is worth studying even if you never migrate.
+Save · ↗Share
Scraping
Crawlee 4.0 dropped overnight: JS-heavy sites and CAPTCHAs handled in one call
Stealth-mode rendering is now the default, with engineers reporting failure rates on real-world sites dropping from around 15% to under 2%. It's a drop-in replacement for the closed-source scraping services you'd otherwise pay for, and the release notes read like a checklist of everything that made 3.x flaky.
crawlee.dev
Crawlee 4.0: stealth rendering, CAPTCHA handling, and a rewritten browser pool
Crawlee 4.0 ships stealth-mode rendering by default, a rewritten browser pool, and first-class CAPTCHA handling. Migration from 3.x is one config change for most…
Read the article →
🎯 Why it mattersTrawly's scraping is X and Exa today, but the day you widen the net, this is the tool that lets one person crawl what used to need a team. The failure-rate drop is the headline; the checklist of 3.x pain it closes is the part to skim now so you know what to reach for later.
+Save · ↗Share
Solo GTM
A solo SaaS founder published her full 12-month GTM dashboard
$0 to $48K MRR over 12 months. $32K total ad spend, $0 on content. 98% of revenue came from one channel she didn't expect at month 2. The chart shows the moment that channel started compounding, month 4, and what she was doing differently in the four weeks before it.
Natalia Reva
@nataliareva
12 months of building solo. $0 → $48K MRR.
$32K total ad spend. $0 on content.
98% of revenue from one channel I didn't expect at month 2.
Here's the full dashboard, no edits. Charts + cohort tables + the four weeks that changed everything ↓
❤️ 11,438 · 🔖 6,204
View on X →
🎯 Why it mattersSpecific numbers from a real solo founder beat any guru's framework. The month-4 compounding chart is the one to screenshot. The channel that drove 98% of her revenue isn't the one you'd guess; worth checking whether it fits a personalized-AI product before your own push past beta.
+Save · ↗Share
🎯 Deal Flow
AI Sales
Half of mid-market B2B teams now run an AI sales rep in production, and pipeline is up 30%
A survey of 600 revenue teams found 48% have an autonomous AI SDR sending outbound today, not piloting. Teams running them report 30% more qualified pipeline per human rep. The catch buried in the data: reply rates are flat, so the lift is entirely volume. The thread argues the winners will be the ones who fix reply quality before every inbox is 90% agent-written.
C
Clara Field
@clarasells
New RevOps benchmark just dropped and it's wild.
48% of mid-market B2B teams already have an AI SDR sending outbound in PRODUCTION. Not piloting. Live.
Pipeline per human rep: +30%.
Reply rate: flat.
The lift is pure volume. Whoever fixes reply quality first wins the next 18 months. Data ↓
❤️ 6,910 · 🔖 4,204
View on X →
🎯 Why it mattersYou're about to run outbound for Trawly's push past beta. If half your buyers' peers already have an AI SDR, a generic sequence is invisible in the pile. The reply-quality gap this thread names is your opening: a human, specific first line still cuts through. Design your outreach around that before you scale it.
+Save · ↗Share
Buyer Behavior
The new enterprise buyer runs your product in prod before they'll take a demo call
A go-to-market analyst mapped 200 recent B2B deals and found the motion has inverted: technical buyers self-serve, deploy against a real workload, and only book a call once they've already decided. Marketing that gates the product behind a form now loses to competitors who let buyers straight in. The piece argues 'time to first value' has quietly replaced the demo as the top of the funnel.
reforge.com · 2d ago
Product-Led Sales: Why Buyers Deploy Before They'll Talk to You
Across 200 recent B2B deals, the buying motion has flipped. The demo is no longer the top of the funnel. Time to first value is. Here is what that changes…
Read the article →
🎯 Why it mattersThis is the exact bet Trawly's example digest makes: let people read a real edition before they ever sign up. The analyst's data says that instinct is now the winning motion, not a nice-to-have. Your version of 'time to first value' is simple: did they open one morning and feel seen.
+Save · ↗Share
🧬 Biotech Disruption
DTC Health
Oura is piloting at-home blood panels that sync to the ring, and the waitlist hit 40,000 in two days
The wearable maker is testing a mail-in blood panel that maps 50 biomarkers against the ring's sleep and recovery data, closing the loop between what you wear and what's in your blood. Early-access spots filled in 48 hours. The interesting move is distribution: they already own the daily habit and the trust, so a lab test is a cross-sell, not a cold start. Regulatory clearance is the open question the piece flags.
statnews.com · 3d ago
Oura's next act: at-home blood testing that plugs into the ring you already wear
Oura is quietly piloting mail-in blood panels that map against ring data. A 40,000-person waitlist filled in two days. The bet is distribution, not diagnostics…
Read the article →
🎯 Why it mattersThe structural lesson maps straight onto your genetic-testing idea: whoever owns the daily relationship can add the lab test cheaply, while a standalone testing brand pays full price for trust. That is the moat question you keep circling. Oura is showing what the wearable-plus-lab bundle looks like in market.
+Save · ↗Share
Precision Medicine
A Nature study split Alzheimer's into five molecular subtypes, each responding to a different drug
Researchers profiled spinal-fluid proteins from over 1,000 patients and found what looks like one disease is at least five, each with a distinct biology and, in early data, a distinct drug response. It reframes decades of failed one-size trials as trials run on mixed populations. The implication for the whole field: diagnosis has to get molecular before treatment can get precise.
sciencealert.com · 4d ago
Alzheimer's May Be Five Diseases, Not One, And That Could Explain Decades of Failed Drugs
Profiling spinal-fluid proteins from more than 1,000 patients, researchers found five molecular subtypes of Alzheimer's, each with its own biology and drug response…
Read the article →
🎯 Why it mattersThe through-line from consumer genomics to this study is one bet: value lives in resolution, telling apart things that used to look identical. If your testing venture sells anything, it sells that resolution. Keep this piece too as a clean way to explain 'precision medicine' to a non-scientist, which you'll need in a pitch.
+Save · ↗Share
🌱 Raising Thinkers
Parenting
Five mental models to teach kids before they're ten, with dinner-table exercises for each
Inversion ('what would make this fail?'). Second-order thinking ('then what?'). The map is not the territory. Confirmation bias, in age-appropriate language. Probabilistic thinking. Each comes with a 2-minute exercise calibrated by age. The piece is short; the exercises are what you screenshot.
Dr. Laura McCann
@drlauramccann
5 mental models to teach kids before they're 10.
1. Inversion ("what would make this fail?")
2. Second-order thinking ("then what?")
3. Map ≠ territory
4. Confirmation bias (in kid words)
5. Probabilistic thinking
A 2-minute dinner-table exercise for each. Thread ↓
❤️ 8,132 · 🔖 22,048
View on X →
🎯 Why it mattersTwo boys, one at an age where 'then what?' actually lands. Second-order thinking is the one to try at dinner tonight: ask what happens after the thing they want, then what happens after that. The others keep. This is the section that has nothing to do with the build, and it's the one you said you never skip.
+Save · ↗Share
🧵 Thread of the Day
“I sold my company for $80M after 9 months. Here's the math nobody talks about.”
A 200-tweet thread that went viral overnight. Full P&L, the acquirer's rationale (which the founder genuinely didn't expect), and his regrets. The pricing structure is the surprising part: flat $19/mo with three soft-paywall tiers.
founderpostmortem
@founderpostmortem
9 months from launch to $80M acquisition.
Full P&L below. The acquirer's rationale (which I genuinely did not expect). My biggest regret.
The pricing structure is the part you won't predict: flat $19/mo with three soft paywalls.
Long thread. Read it slow. ↓
❤️ 47,182 · 🔖 38,410
View on X →
+Save · ↗Share
🃏 Wildcard
Today's outside-your-orbit thing.
A single spot in Venezuela has been hit by lightning every night for 500 years, and it once stopped Sir Francis Drake
Lake Maracaibo's southern basin is the most lightning-dense location on Earth: 1.6 million strikes a year, up to 28 a minute at peak, visible from 400km away. Sailors in the 1500s navigated by “el faro de Maracaibo.” In 1595 it foiled Sir Francis Drake's planned night attack on the city; the lightning revealed his ships' positions before he could land. Locals said for centuries it was the souls of conquistadors. A 2024 atmospheric paper finally pinpointed the real cause: methane bubbling up from the lake bed, combined with a freak mountain-air pattern that pulls moisture upward at exactly the right angle. Tell the boys at breakfast: there is a lake where the lightning never stops, and for 400 years nobody knew why.
via atlasobscura.com
+Save · ↗Share
Editor's Corner
The Batch API cut is the one to act on this week: your compile step was built around cost ceilings that no longer exist, so the Opus-everywhere question is open again. Notice what Deal Flow and Biotech were circling from different rooms, whoever owns the daily relationship gets to add the next thing cheaply. That is the same moat question sitting under both Trawly and your testing venture, and it is worth an hour with a coffee and no laptop. The lightning lake is the one to think about while you fall asleep.
— Trawly 🐟
Trawly
That's about 76 minutes of doomscrolling saved today.
↑ used to be me!
Go do something good with it.
Give David BETA feedback →
Trawly is in beta, your reply shapes what comes next.
The most important thing in my morning, picked from everything Trawly read overnight. One story, not a feed. Some of my best product calls started right here.
— D
This little gold note is the whole point. It takes the story and tells me why it matters to MY week: what I'm building, what's blocking me. Without it, this is just news.
— D
These sections are mine. Deal Flow and Biotech weren't here two months ago; I added them as my focus shifted and Trawly followed. Yours would be built from your profile, and would move as you do.
— D
My favourite part. Always outside everything I asked for, on purpose. This is Trawly's taste, not mine. The reason I open it every single day.
— D
Trawly's own voice at the end. It's watched what I react to, and it tells me what to actually DO with the morning. Most days I don't argue.
— D
That was David's morning. What would yours look like?
