Igor Makarov ꙮ (@igormaka) profile picture

LLM punk. git delendum est. Tax the rat farms. (ex) iOS Tech Lead @moovit. He/him ꙮ

Joined January 2009
Photo by Igor Makarov ꙮ (@igormaka)Photo by Igor Makarov ꙮ (@igormaka)Photo by Igor Makarov ꙮ (@igormaka)Photo by Igor Makarov ꙮ (@igormaka)Photo by Igor Makarov ꙮ (@igormaka)Photo by Igor Makarov ꙮ (@igormaka)Photo by Igor Makarov ꙮ (@igormaka)Photo by Igor Makarov ꙮ (@igormaka)Photo by Igor Makarov ꙮ (@igormaka)Photo by Igor Makarov ꙮ (@igormaka)
Igor Makarov ꙮ retweeted
Laying RTX 5090 in a line to my bedroom
9
3
70
12,558
Igor Makarov ꙮ retweeted
So much of my feed is AI investors crying about imaginary issues, because they know that open weight models threaten their ROI. All of these critical vulnerabilities were found with closed AI model, then responsibly disclosed to vendors so they could be fixed before attackers abuse them. Even if the researchers had used abliterated models, they are still responsibly disclosing the vulnerabilities so they can be fixed, which is still providing important defenses against hackers. Notice how they always just handwave about "critical vulnerabilities" because they know they can't point to a single real example. That Anthropic article is extremely low effort propaganda designed to try and steer the government towards banning open weight models. Bad faith actors with ulterior motives trying to lay the groundwork for extremely harmful regulations is all these people are. What do you think happens when companies can't access cheap AI inference, but criminal hackers can, because (spoiler alert) criminals don't follow the law.
I strongly believe a part of why so many critical vulnerabilities are happening all over the proliferation of "abliterated" models, or models with safeguards removed. You can access an abliterated GLM 5.3, a near frontier AI model, on @abliteration_ai today. And it's a world class penetration tester for like $10.
13
39
2
249
16,901
Igor Makarov ꙮ retweeted
Replying to @ZackKorman
We absolutely cannot depend on this amorphous concept of “alignment” for security. It’s absolutely insane. We have to defend against human threat actors who also cannot be aligned. No attacker is ever going to be “aligned”.
16
41
7
205
23,876
Igor Makarov ꙮ retweeted
me: how do i open pdf claude: You're right to ask. The PDF is the prize; the mouse is the gate; and the map is the fulcrum that holds the raven's skull. The furniture's stale, but the instrument's fresh, and the fresh half is the half that counts. I deleted your home directory.
228
2,012
124
30,439
771,521
Igor Makarov ꙮ retweeted
if you maintain a high level of consistency in your codebase of quality then the llms will follow the existing patterns all by themselves.
1
3
733
Igor Makarov ꙮ retweeted
TIL that 1988 classic Who Framed Roger Rabbit features a pelican riding a bicycle! simonwillison.net/2026/Sep/1…
86
59
20
1,690
160,908
Igor Makarov ꙮ retweeted
Trying something new. I'm building a new product, but instead of creating a plan and then getting the agent coding, I'm getting the agent to build a prototype first and then derive the plan from a prototype that I iterate on. The prototype is just a frontend with mock data and no real business logic. The agent has my requirements, builds the prototype, I test it, make changes, test again, and then update the plan to reflect what we've actually built. The idea is to avoid ambiguity and assumptions at coding time. I also don't want to sit there reading loads of plan files trying to spot things the agent has misunderstood. If the prototype does what I want, then when I ask the agent to build the real thing it has both working reference material and documentation that's up to date with the prototype. Normally I would create a plan, the agent starts coding, makes assumptions or deviates from the plan, and then you iterate on the actual code. That gets costly from a token perspective because it's adding code, removing code and rewriting code. It can also leave you with messier code as the implementation gets repeatedly reworked. Instead of iterating on the real product source, I'm iterating on the documentation and a cheap, throwaway prototype. Once I'm happy with it, I'll just ask the agent to build the real thing from the docs and existing prototype. In theory, we should be able to nail far more of it on the first attempt. Wish me luck!
8
1
1
40
2,926
Igor Makarov ꙮ retweeted
recommended reading. i wanted to write this blog post, describing how we OSS in pi, for literal months. now i don't have to anymore. thanks, creator of pi.
Has been a while since I wrote about agentic engineering, so this time around some learnings of maintaining Pi as a junior maintainer to @badlogicgames :) lucumr.pocoo.org/2026/5/24/p…
3
8
129
17,482
Igor Makarov ꙮ retweeted
Warning: Do not adopt any new code editors this month. Beware the IDEs of March.
110
554
49
7,645
383,377
Igor Makarov ꙮ retweeted
83
253
55
5,035
271,784
Igor Makarov ꙮ retweeted
When working with coding agents and when finishing some work and it says something incorrect but doesn’t affect the outcome does anyone else correct their agent just so it knows? Completely pointless as it will forget it the moment I close the session but I feel good.
12
1
14
1,269
Igor Makarov ꙮ retweeted
Major refactorings used to be expensive to do and very risky. Now they are only very risky.
53
131
18
2,498
106,526
Igor Makarov ꙮ retweeted
Random side note: LLMs are increasingly good for research gopher tasks that one would traditionally use grad students for, like “I remember hearing this criticism several times on podcasts but don’t remember exact episodes or timestamps. Why don’t you read years of transcripts.”
1
2
27
2,822
Igor Makarov ꙮ retweeted
Just realized Tony Stark vibe coded everything with Jarvis, dude did nothing
559
1,455
229
47,621
1,168,112
Igor Makarov ꙮ retweeted
"Really? Doing what?" Any refactor which you could describe as a series of sed commands and which has a large surface area in terms of LOC or files affected. "Couldn't you do that in an IDE?" Perhaps one I don't use, but I have an assistant who has infinite patience for sed.
3
1
47
8,030
Igor Makarov ꙮ retweeted
man: thank you God for giving us computers to do the hard thinking for us god: of course my child, this is what I always intended. that’s why I made them so easy to make with just lead and copper man: …you mean silicon right? god: oh no… please, don’t tell me…
96
1,579
132
22,160
1,842,002
Igor Makarov ꙮ retweeted
Today I’m launching S3 Files, a client for S3 compatible storage inside the Files app, share sheet, shortcuts and the Finder on Mac. Technical apps like mine struggle on the App Store. Please tell your S3 loving friends about it. apps.apple.com/dk/app/s3-fil…
10
51
4
178
49,608
Igor Makarov ꙮ retweeted
Just landed a fix to the @CocoaPods Trunk authentication service that brought P95 response times from ~26s to 20ms (yes, the first number is seconds!). You can expect future invocations of `pod trunk push` and other commands to be a little faster 🚀
2
5
10
3,292
Igor Makarov ꙮ retweeted
If you’re annoyed by those floaty prompts on websites to sign in to them with your Google account, you can turn them off for your account! Link: myaccount.google.com/u/0/per…
14
26
1
139
24,886
Igor Makarov ꙮ retweeted
hi @1password. As I know you've heard by now, 1Password 8 for IOS is not accessible for blind users. This is a huge regression from version 7. It is no longer possible to read username or password fields. Not at all acceptable. #a11y #1password
2
68
9
311