Anton in Tech

Navigating the world of digital payments and startups 🌍💳 Always curious

[Chrome @0xchromium]

Epic Games CEO says it took 30 years and tools like Nanite and Lumen to make building a world even possible, now Fable 5 does the same job for one person in a day

the man behind Fortnite and Unreal spent his career giving artists tools so a world didn't depend on a programmer typing every piece of it by hand

and for one kind of world, a book's world, that work is already done, and you don't even need to be the person who did it

> give a model the full story, get back the world brief, places, look, scale
> a 2nd model turns that brief into a playable space you can move through
> you never open an engine and you never write a line, you just direct
> it runs in a browser, so a reader steps inside without installing anything

the people who live inside these stories have wanted to walk into them forever and never had a way in, let alone a way to own it

30 years of work to make building a world possible at all, now compressed into one prompt and an afternoon

[CyrilXBT @cyrilXBT]

140K GitHub stars. Three months.

Most people still copy-pasting into ChatGPT every morning.

Hermes plus Obsidian plus NotebookLM equals a second brain that compounds forever.

Writes its own skills. Maps its own knowledge. Remembers everything you taught it.

One setup. Runs locally. Never starts from zero again.

Save this before you explain yourself to an AI one more time.

Bookmark this now.

[Blum @Blum_OG]

Boris Cherny put the agent shift plainly:

"it actually uses tools. It acts in the world."

that’s where Claude agents start getting serious

how to build one that survives real work:

1. write the task file
- what it does
- when it runs
- what inputs count
- where output goes
- what it must avoid

2. add memory
- past outputs
- failures and fixes
- edge cases
- patterns worth remembering

3. build recovery paths
- retry tool failures
- save partial work
- alert a human only when the agent is truly stuck

4. add quality checks
- define what good output looks like
- check each result before it ships
- retry with specific corrections
- save failures instead of hiding them

> start with one recurring workflow
> run it for a week
> fix what breaks

the deeper build pattern is in the article below

[Gabe @TheGabeCollins]

The market for compute resembles the basis issue in the U.S. natural gas market. Differing chip capabilities/differing basis conditions (storage/transport capacity + weather). Good work @TheStalwart @tracyalloway!

https://t.co/1fBq3cuEvy

[David Weisburd 🚀 @DWeisburd]

Capital itself is becoming a moat in venture.

When a company is raising $25B, the round cannot be built from a long tail of small checks. The founder needs investors who can move quickly, underwrite the company, and write $1B to $3B checks without blowing up their own portfolio construction.

That changes the competitive landscape.

@robstephens_ explains that this dynamic creates "such an opportunity for the big tier one VCs" because they are one of the only groups that can keep playing this game at the scale founders now require.

It is one of the reasons the largest firms and sovereigns keep extending their lead. In a world of mega-rounds, access is no longer just about brand or relationship. It is also about having enough capital to be relevant.

[Fokki @0x_fokki]

🚨an anime studio spends $180,000 per episode

a 22-year-old spent $57 last month and made $14,600
he built it in 3 weeks.

> Claude maps every script: 10 minutes
> Midjourney designs every frame: 20 minutes
> Runway turns storyboards into motion: 15 minutes
> ElevenLabs delivers every voice: 10 minutes
> Suno produces the score: 5 minutes
> Make releases on schedule: 0 minutes

60 minutes per episode. 4 episodes/week.

the full pipeline is in the article above

[Blum @Blum_OG]

"I don't even talk to Claude. I have a Claude that's talking to my quads"

that line from Boris Cherny gets close to the new job of the engineer

the control loop is what lets coding agents survive stale context, failed tests, partial plans, and premature victory laps

the article's operating model is clean:

1. write the goal as a contract

say the desired end state, the evidence required, the constraints, and the turn or time budget

2. pick the verifier before the run starts

tests, type checks, lint, integration runs, browser screenshots, benchmark scripts, held-out evals

the agent's confidence doesn't count

3. split execution from judgment

one agent builds

another agent, script, test suite, benchmark, or vision review decides whether the work is good enough

4. keep the outer loop alive

wake up, inspect progress, run checks, compare against the goal, send the next instruction, repeat

5. leave artifacts

logs, screenshots, loss curves, benchmark scores, changed files, cost estimates, recent decisions

markdown can store the evidence
an HTML view can help the human supervise several runs

6. mine the sessions

when the same mistake appears 3 times, promote it into project instructions

the new skill for engineers is writing the conditions for autonomy

> a good agent can keep typing
> a good system knows when to make it keep trying

[FHILY👑 @Oluwaphilemon1]

Here is the prompt :
https://t.co/IVC5g9OVHo

[rvaniaaa @rvaniaaaa]

Most people think AI design looks generic because the models aren’t good enough.

That’s not the problem.

The problem is that most people generate screens before they generate rules.

A button.

A dashboard.

A settings page.

Then another screen.

Then another.

Five prompts later, it looks like five different products.

The best designers don’t design screens first.

They design systems.

The screens are just a side effect.

[Trackmind @0xTrackmind]

Jensen Huang, CEO of Nvidia:

"Every engineer is going to have and manage hundreds of agents."

But most people have not even configured one that survives past day 9

You can't manage 100 agents if your first one forgets everything after every session

No CS program teaches this

Not harness engineering
Not agent memory architecture
Not systems that survive production
Not workflows that don't burn your API budget doing nothing

Most people building agents right now are building demos

They break on day 9
They forget every session
They cost money and deliver nothing

One builder mapped the whole thing out

Free
Step by step
No gatekeeping

This is the complete guide to building AI agents that actually work in 2026

Bookmark it for the weekend

[NeilXbt @neil_xbt]

The difference between a tool and an infrastructure is the same as the difference between a taxi and a car you own!

The taxi gets you where you're going.

The car learns your routes, keeps your music, and becomes more useful the more you use it.

Most AI setups are sophisticated taxis.
Useful. Not compounding.

Three properties define infrastructure. Persistence that remembers yesterday, last week, and six months ago.

Autonomy that runs scheduled processes without you initiating them. Compounding that grows more valuable every week as accumulated context improves every output.

Every cycle produces outputs that improve the next cycle.

The infrastructure that runs in the background while you focus on everything else.

This complete setup guide covers every component, every workflow, every connection.

Follow @neil_xbt for more personal AI infrastructure builds.

[FHILY👑 @Oluwaphilemon1]

Are you tired of constantly paying Claude to reread the same files, only to forget everything once you upload.

Traditional Software engineering solved a variant of this a long time ago: reduce your time-complexity by storing things in persistent storage. Memory startups try to solve this, but they don't have the structure to truly build deep memory.

Stateful swarms is an attempt to solve this problem.

[Mike Zaccardi, CFA, CMT 🍖 @MikeZaccardi]

GS: The AI capex boom has lifted semiconductor profitability but will be an increasing headwind to mega-cap tech ROE going forward.

Semiconductor margins have climbed to a record high $SMH

[NeilXbt @neil_xbt]

A virtual assistant charges $2,000 a month to do what these Claude workflows run automatically for free!

Most people have no idea the workflows exist.

Every morning, you check the same sources for trends. Rewrite the same post for three platforms. Pull the same numbers into the same report. Send the same follow-up email after every call.

None of that is thinking. It is repetition. And repetition is exactly what automation takes off your plate.

The chat answers. The automation acts.

→ morning briefing: calendar, emails that need you, overnight trends — one message before coffee
→ trend workflow: 5-7 post ideas with angles written, delivered before you open your phone
→ repurpose workflow: one article becomes a thread, carousel, and video script in 90 seconds
→ client report: assembled, formatted, and summarized before you touch the numbers
→ follow-up workflow: draft waiting after every meeting with next steps already pulled

Once you see how much of your week is you being a slow trigger for work that does not need you, you cannot unsee it.

The complete build guide. Every workflow by role. Your first one shipped in thirty minutes.

Follow @neil_xbt for more Claude workflow builds that show you what automation actually looks like when someone builds it properly.

[RetroChainer @RetroChainer]

Claude Fable 5 shipped today. The model isn't the lecture. The metric is.

Anthropic tracks one number for Claude: time horizon how long an agent stays coherent on a single goal before drifting.

Claude Opus 4 could build a feature. Six months ago: an overnight run.

Fable 5: days on one goal, per the keynote.

The practical part nobody writes down:

Keep your dead prototypes. The ones that almost worked. Re-run them on every Claude release.

They're your private eval for the exponential.

Most teams test the new Claude on tasks that already worked. That measures nothing.

When a broken prototype flips to working that's the signal. Ship that week.

The capability didn't arrive. It was scheduled. Your prototypes are the calendar.

[George Noble @gnoble79]

POV: Investors watching their SpaceX stock after buying the largest loss-making IPO in history

[Moses Kagan @moseskagan]

~Infinitely better than the standard LA new construction apartment building

Kudos

[Sarbjeet Johal @sarbjeetjohal]

...and they did do that... Claude Fable 5 and Claude Mythos 5.

Here is the link: https://t.co/IB0ryZ0KIF

#ClaudeFable5 #ClaudeMythos5 #Fable5 #Mythos5

[Fokki @0x_fokki]

🚨 a B2B intelligence firm charges $50,000/year for one market report

a 22-year-old spent $60 last month and made $8,400 from the same data

> Claude Code scraped every competitor in his niche: 45 minutes
> Claude Code mapped pricing across 12 marketplaces: 30 minutes
> Claude Code clustered 50,000 data points into demand gaps: 40 minutes
> Claude Code packaged the gaps as alerts: 20 minutes
> Make sends them to subscribers: 0 minutes

total: 2 hours 15 minutes, $60/month

the data was free. the structure sells

all five database types in the article above

[Daniel A. Saedi (DataManDan) @TheRealDanSaedi]

Every sales and marketing tool now promises AI will improve your results

but they don't tell you how it will cost thousands of dollars and hours of time to onboard

We built a system with @OpenAI for the part everyone hides: onboarding

And today we're giving it away for free