Vlad Haltakov
banner
haltakov.net
Vlad Haltakov
@haltakov.net
130 followers 330 following 830 posts
Building products in the age of AI. Automating everything. Ex-VP Eng (cybersecurity), ex-BMW (self-driving).
Posts Media Videos Starter Packs
Ideal setup for coding with AI right now is to work on 2 things in parallel - a simple one and a hard one.

The hard problem requires more concentration, more research and longer coding sessions.

While the AI is coding/researching I switch to the easy problem.
My day as a founder with 2 kids

7:15 - wake up, coffee
7:45 - drive kid to school
8:30 - coffee with wife and baby
9:00 - work
11:30 - walk with wife and baby
12:30 - lunch
13:00 - work
15:45 - pick kid from school
16:30 - family time
21:00 - kids in bed, work
00:00 - go to bed
We had a quick planning session with my friend with whom I'm working on Simple Photo Gallery to outline what we need to do before launch of the web app.

We used the best planning tool - a text file 😅

Are we gonna make it by the end of this week?
Cheetah is not as good as Sonnet 4.5, but it is so much faster that I prefer to use and and correct it if needed. Still much faster than Sonnet 4.5.

GPT-5-Codex on the other hand is so slow that I don't really care if it is better or not
The Simple Photo Gallery web app is coming close to being ready for launch.

Improved the gallery creation process a lot today, while my partner is working on a great asset manager tool.

We want to launch a simple version first and then start adding some really cool AI features
The number of edge cases. For example, the user subscribed to a yearly Pro plan, then downgrades to a monthly Basic (but it starts when the yearly ends) and then upgrades to a monthly Advanced (which replaces the Basic one). Then the current plan is Pro, but the future one is Advanced
Handling upgrade and downgrade of SaaS plans using Stripe is quite tricky. The Stripe API helps a lot, but still there is so much to take care of...
The open-source version of Simple Photo Gallery is quite solid now. I'm actively using it myself.

Switching focus to the SaaS version now
Cheetah knows Breaking Bad
So, it seems creativity is still an area LLMs struggle with. They may be able to produce impressive responses, but without much variaty in them.
For comparison, these are my results, so I'm officially more creative than the SotA LLMs 😅

That being said, there are much more creative people than me, like for example @waitbutwhy.

https://x.com/waitbutwhy/status/1426570321109336069
I gave each LLM the exct same instructions that a human gets and asked them to write down the 10 words.

Funny thing is that Grok messed it up and couldn't even come up with 10 unique words, repeating 2 of the them, but since the score only considers your 7 best, it didn't show.

👇
Recently @karpathy talked about how unvaried LLM responses are because of distribution collapse.

So, I wondered how creative LLMs will be on the Divergent Association Task - come up with 10 words that are as different as possible from each other.

Spoiler alert - they suck 👇
ChatGPT thinks I'm suicidal...
Not much to do at the moment in my consulting job, so I'm spending all my time on my projects. Making a lot of progress with Simple Photo Gallery!
I'm back in the flow using AI agents!

The new, mysterious Cheetah model in Cursor is so fast that it changes my whole workflow

Instead of submiting a task and switching to another window to do something else, I can wait because the changes are almost immediate. Back in the flow
The Hetzner SX65 is about $120 and has 4 x 22 TB SATA HDD and 2x 1 TB.

Assuming we only use the HDDs for storage and a RAID 1 configuration for data redundancy, this is more than 2x cheaper than the cheapest object storage!

Sure it requires some setup, but...
Everybody knows @Hetzner_Online is very cheap for compute, but it is also super competitive for storage!

See these prices for TB/m
AWS: $23
GCP: $20
DO: $20
Azure: $18
Cloudflare: $15
Wasabi: $7
Backblaze: $6
Hetzner Object Storage: $6

And then the beast Hetzner SX65 - $2.63!
Thanks to @fiveoutofnine for collecting and curating all images.

https://nyc-friends.vercel.app/
I love AI and AI tools but an always on AI device, listening to everything all the time is just creepy.

It is interesting to see how many people thing the same and are defacing the ads for friend .com in NYC.

https://simple.photo/gallery/friend-nyc/
Browser agents are lazy bastards

Ask them to do some repetitive work and they just hallucinate the answer. Gemini does it and OpenAI Atlas does it. They were supposed to save us time, not "simulate" the work!

@sama @OfficialLoganK make these GPUs go brrrr please!
OpenAI Atlas cannot download images...
Oh wow, OpenAI Atlas is not blocked in the EU...
How can I misspell my own name twice in a simple repo 🤦‍♂️