xvldslv

Founder @polygram_trade

Prev: Founder and exCEO @WagmiHub (exit) & @adly_tech (exit).

[xvldslv @xvldslv]

my coding workflow right now and honestly it's hard to go back

codex handles all the backend and heavy lifting

> database migrations
> api endpoints
> complex refactors across multiple files
> anything that takes more than 20 minutes of focused work

for complex tasks i use exec plans — basically a detailed design doc that codex follows step by step without asking you questions every 30 seconds

it researches the codebase, writes the full plan, implements everything, updates the plan as it goes, and commits along the way

you review the plan once before it starts, then walk away

link to exec plans if you haven't seen this yet: https://t.co/BnCVfWplGe

claude handles everything else

> frontend components and ui tweaks
> quick research ("how does this api work")
> small bug fixes
> anything where i need a fast back-and-forth conversation

but the real unlock is combining them

i connect codex as an mcp server to claude, switch claude to plan mode, and tell it:

"give me the implementation plan, but only present it to me after codex has reviewed and approved it"

so claude drafts the architecture, codex pressure-tests it against the actual codebase, and what lands on my screen is a plan that's already been validated by the agent that's going to build it

two agents checking each other's work before i even read the first line

this is the setup. steal it.

[xvldslv @xvldslv]

Nvidia's new chip generates AI responses 15x faster. Here's how.

They bought Groq — a startup that replaced slow memory with fast memory.

Think of it like swapping a hard drive for RAM. Same idea, just at chip level.

The problem? Fast memory is tiny and expensive.

A normal AI chip has 63 megabytes of it. Almost nothing.

Nvidia's fix: pack 128 gigabytes into one server rack.

The result?

400 tokens per second on a model the size of GPT.

That's faster than you can read.

They split the work in two:

- Raw math → runs on the fast memory
- Everything else → stays on normal Nvidia chips

Best of both. No tradeoff.

Oh and they're sending their next-gen chip to space.

No new gaming GPUs though.

[NVIDIA Newsroom @nvidianewsroom]

NVIDIA Vera Rubin is opening the next frontier of AI.

#NVIDIAGTC news: The Vera Rubin platform’s seven chips are now in full production to scale the world’s largest AI factories.

Vera CPU, Rubin GPU, NVLink 6, ConnectX-9, BlueField-4, Spectrum-6 and Groq 3 work together as one AI supercomputer powering every phase of AI. https://t.co/GqYcF1sfRg

[xvldslv @xvldslv]

@MartinezOnChain yeah, probably. but from a user's perspective, its basically the same thing. that's exactly why prediction markets are getting so much traction. peoples urge to gamble has been around for hundreds of years

[xvldslv @xvldslv]

Paul Graham (YCombinator Co-Founder) explains what actually matters when you want to be a founder.

This is the best advice I've seen on the topic:

[xvldslv @xvldslv]

Cursor built a benchmark from real engineer sessions. here's which models actually win.

they use a hybrid online + offline benchmarking process:

offline - internal benchmark built from actual Cursor engineer work sessions. average task: 352 lines changed across ~8 files. way harder than any public benchmark.

the prompts are short too - engineers don't spell out every detail. but the grading rubric does. just like real work.

online - controlled analysis on live user traffic. catches things offline misses, like when agent output looks correct but devs still reject it in practice.

both evals correlate and rank models the same way:

> GPT-5.4 on top
> Opus 4.6 and GPT-5.2 right below
> Composer 1.5 (Cursor's own model) beats Sonnet 4.5 - and runs faster on Cerebras
> OpenAI models also use fewer tokens per task

surprised to see OpenAI this dominant in a real-world coding eval. but the numbers are there.

Cursor expects most dev tasks to shift to long-horizon agents running autonomously in the cloud within a year. the benchmark will have to evolve with that.

full post: https://t.co/2gUJGbfNwo

[xvldslv @xvldslv]

99.7% of devs are vibecoding wrong. the 0.3% optimize their apps for agents, not just users.

i build everything with AI agents now. here's what actually sticks across projects:

> agent mode: app runs optimized for the agent to self-test. minimal logs, errors crash everything, auth disabled. one command, one request, app shuts down. zero context noise.

> docs/ graph replaces one fat AGENTS.md. lazy-loaded context tree, like OpenAI's Harness Engineering approach. agents maintain the structure themselves.

> clean code matters 10x more now. agents work dramatically faster on well-maintained codebases. tech debt costs you agent performance directly.

> small utility scripts extend agent capabilities. they plug into the docs graph, save context tokens, give agents rails.

stop optimizing your prompts. start optimizing your codebase.

[xvldslv @xvldslv]

@cryppimagic thanks for the service, man! btw how do you usually manage the execution lag when entering these delta-neutral setups across multiple exchanges? or it's not a concern for this arbitrage?

[xvldslv @xvldslv]

@Rewkang it's insane to get 90% success rate from just 5 hours! what sports do you think will be hardest to crack?

[xvldslv @xvldslv]

@hooeem it looks like you had to dig through a ton of staff to map this out.

congratz on the high success with an article!

[xvldslv @xvldslv]

yeah, speaking from experience (having raised a few pre-seed rounds myself), the main priority is always the team, not the product.

like they said in a YC lecture: if your group partners are blown away by your progress at your bi-weekly syncs, then you’re gmi.

the point is, if you move fast enough, the direction matters less. you’ll figure it out as you go

[xvldslv @xvldslv]

https://t.co/6IWR38qhob

[xvldslv @xvldslv]

@varun_mathur Man, when was the last time you touched grass? It’s like 4th release from you in autoresearch field in a couple of days. Great work!

[xvldslv @xvldslv]

@k1rallik Do you have the guide how to set it up?

[xvldslv @xvldslv]

https://t.co/LyMMAalPP5

[xvldslv @xvldslv]

@midascabal right. now tell me what would happen if ai suddenly went down for a week

[xvldslv @xvldslv]

@CodeByNZ same lol

[xvldslv @xvldslv]

@mfranz_on agree with the others, codex definitely has better limits. it's also way better at handling long-running tasks, especially if you set it up in a feedback loop. meanwhile it seems to me that claude just tries to produce as many output tokens as possible haha

[xvldslv @xvldslv]

@SearchForRyan you should definitely check out execplans. it keeps claude on task so it doesn't get drifted, and it will eventually save you tokens as well.

the article is for codex, but the strategy works great for claude too
https://t.co/mPYWsCGuBZ

[xvldslv @xvldslv]

@damianplayer of course he’d say that, it's in his best interest to keep the ai hype train going so they can make some money before the investment capital runs out

[xvldslv @xvldslv]

@birdabo and then news like this will drop

https://t.co/fX1huDSulY

[Polymarket @Polymarket]

BREAKING: Amazon reportedly holds mandatory meeting after “vibe coded” changes trigger major outages.