Vito 🩻

e/acc tech nerd, ex-engineering director @ Snap (Ads & Monetization), ex-Meta. Running the lab at Halo AI

[Vito 🩻 @vitostrokov]

The most idiotic thing you can do is to ask a user to wait before letting them access your website, because “there are people in front of you using our website”

[Vito 🩻 @vitostrokov]

@dhh @satyanadella What’s wrong with Google Docs?

[Vito 🩻 @vitostrokov]

Exactly my point: What are the “Values” applied here. SF/Bay Area should be top right corner above all in the “Expensive but worth it” category, and London should follow right after. Zurich should place more 3/4 top-right.
Every single site in the top right corner is not even close to SF/Bay Area and London in terms of opportunities, quality of life, access to most desirable things.

[Vito 🩻 @vitostrokov]

What is a “Value” for one could be a “DeValue” for another. A better chart will allow a user to select what are the “Values” for them and rerank the cities accordingly.
I imagine SF will be placed number one for someone who is trying to build a startups and raise capital and has a socialization with founders as the highest value for them, this will push most of the currently high value places like Porto, Lisbon, Stockholm, etc down, because they have non of that.

[Vito 🩻 @vitostrokov]

In California a woman that makes let’s say 60K a year will pay less than $500 out of pocket for the procedures described in the video. In most other states the situation will be similar, only “high income” individuals will be asked to pay out of pocket for child birth.

On the flip side EU does not offer most advanced treatments that US has today: a simple example is an ACL surgery. Half of the clinics in the US offer BEAR ACL repair, whereas all surgeons in EU never heard of it and will be happy to graft your hamstrings leaving you with a permanent 20% loss of function of one of your legs.

US Healthcare system is quite broken in terms of providing affordable general healthcare. Yet quite advanced in terms of novel treatments that save lives and preserve health span.

Ideally you want to have access to both if you can.

[Yann LeCun @ylecun]

Only in America.
Tired of winning?

[Vito 🩻 @vitostrokov]

It actually did pretty well, especially considering that “ugly” pages often convert better.

Would be great to see the prompt you used, @MatthewBerman.

During my time at Meta, I was also surprised by how often what UX/UI designers considered “high-quality” design actually converted fewer leads than something they would consider “poor quality.”

Beautiful does not always mean effective.

[Matthew Berman @MatthewBerman]

I gave Grok 4.6, GPT 5.6 Sol, and Fable 5 the same prompt to generate a profile card for a creator and wow Fable did SOOO bad.

[Vito 🩻 @vitostrokov]

Rate Limit / Caching bug @thsottiaux
It seems like if a session is cached (Codex/Model, in my case 5.6Sol), it will keep accepting new requests unless you add a new prompt manually. That probably either breaks the cache or triggers a rate limit check.
However, if you just let your long-running agent run on its own, it never hits the rate limit. At least in my case it has been running for 12+ hours now.

[Vito 🩻 @vitostrokov]

@prioritypasscom and @AmexBusiness are broken.

Most lounges simply refuse to honor your Platinum membership now, citing reserved space for airline priority customers or saying you need to pre-book, while not actually having a booking page.

What still keeps Amex going for me:
- Insane level of consumer protection when you need a refund or dispute a charge.
- The credit hit you can take if you cancel a long-standing Amex account.

If you do not have Amex yet, I would not get it for the travel benefits.

They simply do not pay back nearly as much anymore.

[Vito 🩻 @vitostrokov]

@elonmusk 7 hour flight with @TAP_Portugal to NYC. No internet. At all. Forget about Starlink.

So yes, next time I’m definitely double-checking this before booking.

I can’t leave @claudeai running unattended for 7 hours, guys.

[Vito 🩻 @vitostrokov]

AI in engineering today is only as good as your own skills and experience as a system engineer and architect.

For simple problems, modern coding agents are almost magical.

Call an API. Transform some data. Write a script. Build a landing page. AI can do 85-99% of the work. But the curve changes VERY quickly once you start building real systems. It is still painfully difficult to get a genuinely good complex agentic workflow out of modern vibe-coding agents.

Even a reliable ReAct agent becomes a serious engineering problem once you introduce:

• persistent state
• multiple tools
• async workflows
• retries and failure recovery
• context management
• permissions
• observability
• evaluations
• multiple data sources

The agent can write almost every individual piece of code.
That does NOT mean it can design the system.
Move one level further into multi-tenancy, distributed data processing, security, isolation, migrations, queues, caching, reliability, and infrastructure, and the gap becomes enormous.
You can vibe-code something that looks incredible in a demo in a few hours.
Vibe-coding something that securely serves millions of users without randomly breaking is a completely different sport.
And this is the strange part about AI coding:
It makes experienced engineers MORE valuable, not less.
AI dramatically lowers the cost of implementation while massively increasing the leverage of good architectural decisions.

A great architect with AI can now execute at ridiculous speed.

Someone who cannot recognize a bad abstraction, broken state model, security vulnerability, race condition, scaling problem, or architectural dead end can also generate an enormous amount of code at ridiculous speed.

The bottleneck is moving from: “Can you write the code?” to “Do you actually understand the system you are building?”

AI can increasingly do the first.
The second is still very much engineering.

[Vito 🩻 @vitostrokov]

@glproductions @levelsio TBD --> Great, talented people. Too much politics. Hence everything is slow and people are yanked left and right, mostly leadership is overoptimizing for what is looking good.
GDM --> IDK, @demishassabis just left, it can't be a good sign for them.

[Vito 🩻 @vitostrokov]

@levelsio Well, @levelsio I DMed you a few months back. And I fall into this category. Let’s connect.

[Vito 🩻 @vitostrokov]

@nickscamara_ @firecrawl Does it extract margin comments from PDF and Docx?

[Vito 🩻 @vitostrokov]

Here is what happens next (next 12-18 months):

1. @AnthropicAI makes Fable significantly less constrained.
2. @OpenAI launches GPT-6, likely a Mythos-grade model or better.
Anthropic launches Mythos 5.1, which will be almost unconstrained.
3. @Kimi_Moonshot launches Kimi K3.5, an open-weights Mythos-grade model.
4. @Apple, @nvidia, and @Microsoft launch local hardware capable of running Kimi K3.5 in NVFP4 at home/office.
5. More companies switch to running Kimi and Qwen locally.
6. Anthropic cuts prices and releases Opus as open weights. Meanwhile, OpenAI, with @sama clearly no longer able to “steamroll” anyone, focuses on making ChatGPT and Codex available through as many partner integrations as possible, while monetizing more aggressively through ads.
7. @finkd finally launches Muse Spark 1.2, a Luna-grade model.
Why such a massive gap?
You should ask the Meta AI team about its internal politics.
8. Meta still launches a GPU cloud, but it is focused primarily on large deals with Anthropic and OpenAI.

[Qwen @Alibaba_Qwen]

Meet Qwen3.8-Max: A New Bar for Coding and Cowork.

[Vito 🩻 @vitostrokov]

@ColinGardiner You can request access to our deck here: https://t.co/GAvZjc21AJ

[Vito 🩻 @vitostrokov]

You either want to solve the problem, or you want to make a point about gun control. You really need to choose.

If the goal is to solve the problem, then you should listen to what @PalmerLuckey is trying to tell you:
Split “kids” into meaningful age groups. Identify why children are dying within each group. Build a focused solution for each one and implement it. Then monitor whether the numbers actually change.

You may be surprised to discover that the isolated group of 16-18yo still dies from gun violence even after you banned guns, because the underlying causes were never addressed.

Ask your AI to “red team” this thread. Use incognito mode. And do not let it drag you into the “I am always right” territory.

[Vito 🩻 @vitostrokov]

@OpenAI should introduce a $499 unlimited plan, capped only by the number of simultaneous threads.

For example, 10 concurrent GPT-5.6 Sol threads at normal speed, or 5 threads running at 1.5x speed.

That would push even more developers to switch from Claude to Codex as their primary coding agent, because there would be little reason to maintain two expensive subscriptions.

cc @thsottiaux