utopian tech bro, crunchin numbers. Research Engineer @PAM/HarmixAI
Madrid/London
Joined March 2015
- Tweets70
- Following345
- Followers43
- Likes127
Vitalii Ratushnyi retweeted
The context a long-running agent works over across days easily runs to millions of tokens and the field still has no consensus on what memory should even look like.
@Misreadable is the Lead Research Engineer at HarmixAI working on long-term memory for agents with inputs of 1–10 billion tokens, and a former quant, now based in Madrid.
He'll compress two years of research and production experience into practical recipes for scalable agent memory.
Figured I spent lot of compute on deep research and bet many of you are doing the same. Decided to assemble results in a structured way in the repo, both for human/agent readers. Only important tech and a structured format for easier context navigation. Pls contribute with ideas and star the repo: github.com/MisreadableMind/a…
Just finished with @stateofai
Tbh, the report could be 60 slides without losing quality.
Primary slide: Users trust ChatGPT as much as direct referrals (if not more)
Vitalii Ratushnyi retweeted
We are giving away 40 water-tight contract templates. This would cost >$10K to any agency, startups, or freelancer, to create from scratch.
Retweet this tweet and comment 'Monk' and we'll send them to you.
Vitalii Ratushnyi retweeted
$3 Trillion is currently stuck in unpaid invoices.
An average invoice takes 59 days to clear.
If you're doing >$10M in revenue, getting paid in 30 days instead of 59 days will literally make you millions over 2 years.
Introducing Monk. It collects your money faster and pays for itself in 30 days, guaranteed.
Fast-growing companies like ElevenLabs and Profound rely on Monk.
The Problem: 39% of the $3T stuck is due to two stupid reasons:
1. "Please fix this comma in the invoice and then we'll pay you'" (this happens 2-3 times per transaction)
2. "Sorry the payment reminder got buried in my inbox" (automated emails get ignored)
A payment that should take 2 days, takes 10 days. Your cash on hand is embarrassingly behind "recorded revenue".
To fix this, we raised $4M led by @btv_vc with participation from @gtmfund and @danonanthony.
For our first product, we had to innovate on 3 dimensions:
1. Monk turns signed contracts into invoices with near-perfect precision
- We leverage frontier models to extract key terms from your deals and turn them into invoices.
2. Monk collects payments agentically with a 24% better reply rate than automated emails:
- Tuned to write emails that feel like a human request, not spam. Your invoice is competing with everything in their inbox.
- We know how to find an alternate point-of-contact when someone is OOO, when to reach out, and what tonality yields a higher response rate.
If you had someone on your team whose invoice emails got answered 24% more than others - that would be a big deal. You’d make them the head of revenue collection.
3. Granular visibility into your cashflow:
- Understand your cash position, aging, or expansion/contraction at customer level.
Book a demo at monk.com/book-a-demo-form and we guarantee that Monk will pay for itself within 30 days.
If you've read this far, we are giving away 40 water-tight contract templates. This would cost >$10K to any agency, startups, or freelancer, to create from scratch.
Retweet this and comment 'Monk' and we'll send them to you.
The quest for greatness can be reframed as a search problem.
In a world of information abundance, how do you find good influences? (C) @phokarlsson
@phokarlsson I'm on pt2. Such an amazing piece thinking/writing!
Ref: henrikkarlsson.xyz/p/trainin…
"NEVER MISTAKE MOTION FOR PROGRESS"
Bad Engineers mistake motion for progress.
Good engineers reach for simple solutions first, write code their junior colleagues can maintain.
Source: terriblesoftware.org/2025/06…
NVIDIA’s 2024 Blackwell GPU
uses 105,000 times less energy to generate tokens than its 2014 Kepler predecessor
Source: bondcap.com/reports/tai - page 135
UX and Quality is next big stop for AI Apps.
As an Engineer with X yrs of Experience i can produce substantially better app (tech-wise) using current generation of LLM than somebody who just started.
So basically having taste and skill to communicate UX is a key.
Vitalii Ratushnyi retweeted
AI ASMR is blowing up
This account was started three days ago with just 11 videos…but has racked up 82k followers 👇
Not by bread alone.
"A subsistence farmer from a thousand years ago would look at what many of us do and say we have fake jobs, and think that we are just playing games to entertain ourselves since we have plenty of food and unimaginable luxuries"
Source: blog.samaltman.com/the-gentl…
"You can vibe code at a buttery-smooth 120fps in Zed and generate code 24 hours a day if you're so inclined. But in a world of abundance, the bar should be higher for quality."
Quality is a king!
Src: zed.dev/blog/software-crafts…
Vitalii Ratushnyi retweeted
An attempt to explain (current) ChatGPT versions.
I still run into many, many people who don't know that:
- o3 is the obvious best thing for important/hard things. It is a reasoning model that is much stronger than 4o and if you are using ChatGPT professionally and not using o3 you're ngmi.
- 4o is different from o4. Yes I know lol. 4o is a good "daily driver" for many easy-medium questions. o4 is only available as mini for now, and is not as good as o3, and I'm not super sure why it's out right now.
Example basic "router" in my own personal use:
- Any simple query (e.g. "what foods are high in fiber"?) => 4o (about ~40% of my use)
- Any hard/important enough query where I am willing to wait a bit (e.g. "help me understand this tax thing...") => o3 (about ~40% of my use)
- I am vibe coding (e.g. "change this code so that...") => 4.1 (about ~10% of my use)
- I want to deeply understand one topic - I want GPT to go off for 10 minutes, look at many, many links and summarize a topic for me. (e.g. "help me understand the rise and fall of Luminar"). => Deep Research (about ~10% of my use). Note that Deep Research is not a model version to be picked from the model picker (!!!), it is a toggle inside the Tools. Under the hood it is based on o3, but I believe is not fully equivalent of just asking o3 the same query, but I am not sure.
All of this is only within the ChatGPT universe of models. In practice my use is more complicated because I like to bounce between all of ChatGPT, Claude, Gemini, Grok and Perplexity depending on the task and out of research interest.
