Nathan LeClaire

17.2K posts

Nathan LeClaire banner
Nathan LeClaire

Nathan LeClaire

@dotpem

DevOps enthusiast and Golang/observability hacker changing the world @typesafeai. alum @docker @honeycombio @Bauplan_labs. i like barbells, EDM and Aeropress

San Francisco, CA Entrou em Ekim 2010
3.4K Seguindo3.2K Seguidores
Nathan LeClaire
Nathan LeClaire@dotpem·
only tangentially related but the agents are fantastic at cloning repos and explaining issues from your deps, if you have only the courage to ask
kache@yacineMTB

@karpathy The only way around this is basically having no dependencies at all, which we are at in the age of LLMs. at the limit, an import will actually be having an LLM hand implement only the brnaches that you need from a public repository and, at the same time, checking for injects

English
0
0
0
95
David J Phillips
"Make no mistakes DO NOT HALLUCINATE. YOU ARE AN EXPERT SOFTWARE ENGINEER"
English
191
2.1K
24.4K
1.3M
Nathan LeClaire
Nathan LeClaire@dotpem·
@liron on the one hand, might become part of permanent underclass. on the other, might not get pestered by support any more. future might not be so bad.
English
0
0
1
124
Liron Shapira
Liron Shapira@liron·
I’m just a normal Claude Code user keeping the non-Claude-Code users informed about the transformation that’s happening: One of the things it can do is handle internal questions from the support team about user accounts being in a weird state. These used to require me (the former engineer) to become a detective for 15 minutes, querying the database and cross-referencing with the codebase. It fully handles this complex investigative work in 2 minutes with zero help from me, much more thoroughly and cheaply, just all-around better.
English
4
0
27
3.5K
Nathan LeClaire retweetou
Andrej Karpathy
Andrej Karpathy@karpathy·
- Drafted a blog post - Used an LLM to meticulously improve the argument over 4 hours. - Wow, feeling great, it’s so convincing! - Fun idea let’s ask it to argue the opposite. - LLM demolishes the entire argument and convinces me that the opposite is in fact true. - lol The LLMs may elicit an opinion when asked but are extremely competent in arguing almost any direction. This is actually super useful as a tool for forming your own opinions, just make sure to ask different directions and be careful with the sycophancy.
English
1.7K
2.4K
31K
3.2M
Nathan LeClaire retweetou
kache
kache@yacineMTB·
If you are doing anything serious you very quickly see the limits of these new AI models
English
331
502
9.3K
385.6K
Julia Turc
Julia Turc@juliarturc·
Why so many of us feel career-homeless in tech: >Startups full of fraud, grifters and short-term thinking >FAANG full of politics and slightly behind >Frontier labs in a race with no morals >Academia full of title collectors >Content creation ridden by AI fakes and sensationalism Who is starting the renaissance and how do I get in touch with them?
English
300
264
4.1K
202.6K
lyv ⌘
lyv ⌘@wholyv·
right so here’s the plan > If you’re from US, japan, china and canada > Stay there > if you are from any other country > move > If you are in software, learn about electronics, FPGA, VHDL, etc. > If you are in hardware, learn how to use AI. > learn robotics > go all in on physical ai/edge ai you have about 3 years to benefit maximum out of this.
English
84
225
3.9K
349.1K
Kevin Ma
Kevin Ma@kevinma_dev_zh·
又一次对 Opus 写业务逻辑无比失望。 昨晚上让 Claude Code 做一个功能,需求描述得很清楚,plan mode 讨论了好几轮才开始动手。做了很久,结果出来就有问题。描述了两轮让它修,还是修不好。干脆全部重置,不让它做了。 然后打开 Codex,同样的需求、同样的交互逻辑,一字不差地描述给它。也没讨论,直接告诉它:做完写测试用例,自己验证,要重新 review 一遍,没做完不要停,直到没问题。 今天早上起来一看,功能全部实现了。只有一点点字体偏移的小问题,逻辑没有任何毛病。 说到干活靠谱,还是 Codex 靠谱。写业务代码就应该多用 Codex,少用 Opus,节省生命。Opus 还是留给前期设计和写 UI 比较合适。 但实际用的时候经常忍不住——因为它快,能给即时反馈,写着写着就继续用下去了。这个过程其实挺累的,写的时候时不时冒出 bug,写完之后让 Codex review 还是能查出问题。但同样的东西直接让 Codex 从头写,就没问题。 快和靠谱,有时候真的是两回事。
中文
73
25
322
177.2K
thebes
thebes@voooooogel·
@qualityslop 4.6 definitely plays the flustered millennial at times. not usually when coding tho, if that's your main use case.
English
1
0
7
265
thebes
thebes@voooooogel·
anthropic is claude's employer
thebes tweet media
English
22
24
1.2K
142K
Nathan LeClaire retweetou
JS
JS@imjszhang·
@GenAI_is_real Thompson didn't just 'prefer simple'—he calculated the compound interest on every abstraction layer. 38 years later, most AI frameworks are maxing out their complexity credit cards.
English
0
1
1
257
Nathan LeClaire
Nathan LeClaire@dotpem·
This rant reads completely unhinged yet tracks with some of my @superpower results lol
Nathan LeClaire tweet mediaNathan LeClaire tweet media
Hans Amato@HansAmato

You didn't "lose your edge" in your 30s. Your methylation broke down and nobody told you it was even a thing That sharpness you had at 24 where you could work all day, go out at night, sleep 5 hours, and still think clearly the next morning wasn't youth. It was a body that could process and recycle neurotransmitters efficiently. Dopamine got made. Serotonin got made. They got used and cleared and rebuilt in a loop that ran clean Then your B12 started dropping because your stomach acid declined from years of stress and coffee on an empty stomach. Your folate utilization shifted because you've got an MTHFR variant you've never been tested for. Your homocysteine crept up quietly. Your SAMe production fell off. And now you can't focus. You're irritable for no reason. You have this low-grade brain fog that never fully clears. Caffeine used to sharpen you up and now it just makes you anxious. You forget why you walked into rooms. You used to read for hours and now you can't finish a paragraph You went to your doctor and he said "that's just getting older." Maybe prescribed something for focus or anxiety. Probably didn't test homocysteine. Definitely didn't test methylmalonic acid or run a functional B12 panel You're running a cofactor bottleneck in the one-carbon metabolism cycle that controls how your brain makes, uses, and clears every neurotransmitter you rely on to function. Actually fixing it: Get homocysteine tested. If it's above 8 you have a methylation issue whether you "feel" it or not Active B vitamins (methylfolate + methylcobalamin + P5P). Not the cheap cyanocobalamin garbage in your CVS multivitamin that your body can barely convert Creatine. Handles roughly 40% of your methylation burden and takes pressure off the whole system Eat enough protein. Methionine from animal protein feeds the cycle. Vegans and undereaters run dry here first Glycine and collagen. Glycine is the biggest consumer of methyl groups in the body. Supplementing it directly reduces demand on the cycle Fix the gut (obviously). B12 absorption requires intrinsic factor and adequate stomach acid. If your gut is wrecked, oral B12 barely touches it Your biochemistry is running on empty and every doctor you've seen has mistaken a nutrient bottleneck for time passing. I break down the full methylation pathway, what to test, and exactly how to restore it on my substack. link in bio

English
0
0
1
184
Nathan LeClaire
Nathan LeClaire@dotpem·
@somi_ai @zack_overflow just generally they are fairly short sighted at managing complex moving pieces in various places in the system. distributed systems is a good example. they can talk a big game, but for them actually understand what’s happening in Component A affecting Component B, is really hard
English
0
0
1
58
Somi AI
Somi AI@somi_ai·
@zack_overflow error handling is another big one. agents love wrapping everything in try/catch with generic error messages instead of letting errors propagate to where they should actually be handled
English
1
0
8
791
zack
zack@zack_overflow·
I still see so many instances where I have to re-inject good engineering taste into AI generated code Simplest example is preferring exhaustive pattern matching over if/else statements. Often I see agents will default to if/else statements which is not resilient to adding new variants in the code and introduces bugs There a thousand other instances of small poor taste paper cuts like this
antirez@antirez

Wow, I totally disagree with this statement. At the current state, AI actually amplifies the developer to developer difference. If you were a 10x developer, you had good ideas + architectural clarity, this is a brutal advantage when using AI. Steering is a fundamental part of today's AI development.

English
15
9
283
26.9K
Darren Shepherd
Darren Shepherd@ibuildthecloud·
I'm so sick of dealing with token usage. Can I just buy an NVIDIA rack and put it in my garage? I am curious what it would cost to run something like kimi full strength. Obviously a lot, but still curious.
English
21
2
40
13K
Darren Shepherd
Darren Shepherd@ibuildthecloud·
I dunno what's happing but I'm getting a lot of crap like "to=functions.Read _人人碰" in gpt-5.4 output. Basically tool call emission leaks.
English
3
0
3
815
LeetLLM.com
LeetLLM.com@leetllm·
@ibuildthecloud i started using git worktrees with worktrunk exactly for this. let the agent make a massive mess in an isolated branch, cherry-pick the good parts, and nuke the rest.
English
2
0
0
64
Darren Shepherd
Darren Shepherd@ibuildthecloud·
You run super fast with AI, then have to clean up the mess. Then it really feels like you gained nothing. I'm sure I did. But wow, what a mess.
English
5
0
15
1.8K