Vess Bakalov

Founder of SevOne and (both exited)
ex. CTO at IBM - Networking and AI and SVP at Core42 - Product
Working on the next one / having fun.

[Vess Bakalov @codr_1]

@mfmereu I don't care what game it is - I am not buying a console. Waiting for PC.

[Vess Bakalov @codr_1]

It says 2 days because I turned it on 2pm yesterday. I am posting at 9:30am local time.

[Vess Bakalov @codr_1]

So my GLM 5.3 affair hit a wall. I burned a week's quota in less than a day (or I would have if the 5 hour limits didn't keep kicking in). Anthropic lasts at least 2 days. Codex too, tho less than Claude.
So honestly - the https://t.co/PTy9xakrQ8 - not that good a value.

[Vess Bakalov @codr_1]

@TheBalkanHacker I am still trying to get myself in the habit of talking instead of typing. So hard.

[Vess Bakalov @codr_1]

@TheBalkanHacker If you haven't played Noita - try it out. It is the most glorious pixel art game with amazing physics. It is 2d, but some of the scenes in this video look like 3d versions of some of the monsters.

[Vess Bakalov @codr_1]

@LukasHozda I don't know him, but I really hope he grows out of this pattern.
He has enough followers and he has earned enough celebrity where he needs to project grace and not what feels like insecurity.

[Vess Bakalov @codr_1]

@photonstorm w0w. 0k

[Vess Bakalov @codr_1]

@elonmusk Come to UAE habibi.

[Vess Bakalov @codr_1]

@davidreinig01 @SebAaltonen I once tried to write some high performance python. The fight was - keep arrays in NumPy (C land). Python would happily serialize in and out of NumPy, and that right there ate so much CPU and RAM...

[Vess Bakalov @codr_1]

@guivilelaa_ GLM is slower and not quite so accurate. But if you have to pay API prices...

[Vess Bakalov @codr_1]

@guivilelaa_ there is cost too, you know...

I just posted this. I switched my implementation stage from Claude to GLM 5.3. Sol still does the reviews. Claude still does the planning. This is the last 24 hours (last 4 hours is GLM 5.3)

[Vess Bakalov @codr_1]

Claude is using Claude Code, and GLM is going through OpenCode.
Cluade has done a lot more commits in in the last 24 hours, so I will have to run this again tomorrow.
GLM is slower in general, and clearly so far somewhat less accurate. But not that bad...
700k LOC codebase.

[Vess Bakalov @codr_1]

Today I was running out of tokens. Again.
2 Claude and 2 Codex Max - 85%+
In my Software Factory you can switch which model you use for each stage.
I turned on GLM 5.3 for coding, Opus for planning, Sol for review.
4 hours in - this is a comparison of GLM vs Claude

[Vess Bakalov @codr_1]

I cracked up on the 'loosies'
I know HW is getting super expensive, and I was lucky to buy the my current HW last year around Sept when I moved to Abu Dhabi so I got good prices. But running my workers locally is much better.
One provider I mean to try is AlphaVPS (Bulgarian HQ BTW). Pricing is better. But don't know what the perf will be. Much more traditional managent interface tho. I do use them for some projects my son is doing but those need almost no resources - just a small influx.

[Vess Bakalov @codr_1]

Not trying to be hard on https://t.co/LKbIePdXZC - I landed on it based on @theo's video.
I run workloads across 4 boxes, distributed based on core count - one lane per 4 cores:
- 2 Boats - 8 cores, 16GB RAM
- ea - MinisForum xAI
- Main server - custom, ThreadRipper

host local?

[Vess Bakalov @codr_1]

@MiaAI_lab same

[Vess Bakalov @codr_1]

@VanpeltVentures The reviews were really mixed. Benchmarks look good. Haven't tried it yet. I did have some writing to do - but it had a whiff of cybersecurity - and it flagged it immediately.

[Vess Bakalov @codr_1]

I haven't tried it yet, but my excitement is diminishing by the minute as I scroll X this morning.

[Mia @MiaAI_lab]

For coding and agentic workflows, Fable 5.2 is not leaps much better than Opus 5.5, but it costs twice as much.

[Vess Bakalov @codr_1]

@wojakcodes Gg, but I think your car is trying to tell you something. It really needs a checkup

[Vess Bakalov @codr_1]

@ChShersh @wojakcodes Even my 5090 in glass case is not safe from Chinese models.

[Vess Bakalov @codr_1]

@0xSero I notice it about a week before usually. This time I think they had a failed deploy. Things got bad, then good than bad again in a 2 wk. period