Dexter

975 posts

Dexter banner
Dexter

Dexter

@dexcompiler

courage & leverage

Amsterdam, The Netherlands Joined Eylül 2022
1.4K Following163 Followers
Dexter retweeted
Konrad Kokosa
Konrad Kokosa@konradkokosa·
I've built a full LLM inference engine in C#/.NET 10. From scratch. Not a wrapper - native GGUF loading, BPE tokenizer, attention, KV-cache, SIMD-vectorized CPU kernels, CUDA GPU backend, OpenAI-compatible API. Solo dev, ~2 months, AI-assisted (not vibe-coded!). First preview is out. Check it out for mode details at kokosa.dev/blog/2026/dotl… and dotllm.dev
Konrad Kokosa tweet media
English
45
165
865
59.1K
Dexter
Dexter@dexcompiler·
Clockworks: Time is a dependency, and once you treat it like one, a lot of awkward engineering starts to become much easier to model dexcompiler.com/blog/clockworks
English
0
0
0
19
Dexter retweeted
Google Research
Google Research@GoogleResearch·
Introducing TurboQuant: Our new compression algorithm that reduces LLM key-value cache memory by at least 6x and delivers up to 8x speedup, all with zero accuracy loss, redefining AI efficiency. Read the blog to learn how it achieves these results: goo.gle/4bsq2qI
GIF
English
1K
5.8K
39K
19.2M
Dexter
Dexter@dexcompiler·
@simonw operational complexity is still not in the set of llm-resolved issues
English
0
0
0
5
Simon Willison
Simon Willison@simonw·
I'm not so sure about this. Not all, but a lot of SaaS moats really do rely on an implementation complexity that's rapidly fading Take SAML for example - a classic example of a feature that is such a nightmare to implement that most SaaS startups delay as long as possible and then hire specialists If that implementation time drops from months to days, it's yet another little piece of moat that just got eroded away
François Chollet@fchollet

Cloning any random piece of SaaS is something that could already be done before agentic coding, and the economics of it haven't changed meaningfully. Before, writing the clone would cost 0.5-1% of the valuation of the legacy SaaS company. Now it might be 0.1%. It doesn't make a difference -- if you can pull it off profitably today you could also have done it profitably in the past. The code is a very small part of the process of making such a clone successful, and the reason legacy software has often bad UX is not because code was expensive to write.

English
65
14
399
74.7K
Dexter
Dexter@dexcompiler·
@KooKiz Mostly because agents still lack a world model and to some extent, still operate based on the tunnel vision of the current user story or task in scope. They're utterly unaware of the higher-order effects that a senior dev would implicitly deem undesirable
English
0
0
1
63
Kevin Gosse
Kevin Gosse@KooKiz·
For my professional projects (ReSharper), I limit my AI usage to "explain this code" or "generate this function". I've tried letting an agent work autonomously a few times (Claude Code, with Opus 4.5). Every time I got the same result: the agent completed its task successfully (with some back and forth), but introduced unneeded side-effects, that would have definitely broken stuff if I didn't catch them. Sometimes it's really counter-productive. In one extreme occurrence, I asked the agent to update our leveldb connector, to stop using IntPtr in favor of safehandles, after we caught a memory leak in some unhandled error conditions. The task is simple, but with some subtle changes to consider (IntPtr are never null, handles can be). It completed the task but introduced so many changes in behavior that I had to carefully review every line of code, and it ended up taking me as much time as if I wrote it myself, if not more. Overall, I believe agents right now are at the level of a talented junior developer: fast coder, will complete complex tasks in record time, promised to a bright future. But it lacks that caution that senior developers have, born from the experience of breaking the production too many times. Agents are too task-focused: given instructions, they will complete them successfully, but fail to consider the risks and global impact of the changes they make.
David Fowler@davidfowl

Ultimately we all want to get here, and the code quality will get better, but right now multiple passes over the code are needed to keep the quality high.

English
16
7
191
37.5K
Dexter retweeted
François Chollet
François Chollet@fchollet·
The best way to use AI is an interface to information that lets you deepen and improve your own knowledge and mental models. The worst way to use AI is as a crutch to outsource and forsake your own cognition
English
102
241
2K
89.3K
Dexter retweeted
pashov
pashov@pashov·
🚨Claude Opus 4.6 wrote vulnerable code, leading to a smart contract exploit with $1.78M loss cbETH asset's price was set to $1.12 instead of ~$2,200. The PRs of the project show commits were co-authored by Claude - Is this the first hack of vibe-coded Solidity code?
pashov tweet media
English
316
552
4.4K
1.7M
Dexter retweeted
OpenAI
OpenAI@OpenAI·
Introducing EVMbench—a new benchmark that measures how well AI agents can detect, exploit, and patch high-severity smart contract vulnerabilities. openai.com/index/introduc…
English
1.2K
1.3K
8.8K
2.5M
Dexter
Dexter@dexcompiler·
@ns123abc They did a good job. For non-mid/junior devs, gpt-5.2 is superior to opus 4.5 imo
English
0
0
1
287
NIK
NIK@ns123abc·
BREAKING: OpenAI CEO admits they sacrificed ChatGPT’s creative writing to chase coding Sam Altman: >“I think we just screwed that up” >“We have limited bandwidth” >“We decided to focus on coding” OpenAI is playing catch-up now.
English
238
126
3K
682.3K
Dexter
Dexter@dexcompiler·
@jessfraz Why do they keep publicizing these unpopular plans though? Someone inside openai def understands the bad optics around this The self-sabotage with these guys is on another level
English
0
0
4
800
Jessie Frazelle
Jessie Frazelle@jessfraz·
seems a bit hypocritical that the folks who scraped all the internet to train their models, now want a cut of what it creates, but aren't willing to pay for what they scraped...?
Jessie Frazelle tweet media
English
130
376
2.7K
96.1K
Dexter
Dexter@dexcompiler·
@mitchellh @NapVeg You'll also need another AI to process the session transcripts at some point It's a brave new world
English
1
0
1
375
Mitchell Hashimoto
Mitchell Hashimoto@mitchellh·
@NapVeg I don't really care, I really just want to know that AI was used. Ultimately, I want to see full session transcripts but we don't have enough tool support for that broadly. One day though.
English
7
0
100
6.2K
Mitchell Hashimoto
Mitchell Hashimoto@mitchellh·
Ghostty is getting an updated AI policy. AI assisted PRs are now only allowed for accepted issues. Drive-by AI PRs will be closed without question. Bad AI drivers will be banned from all future contributions. If you're going to use AI, you better be good. github.com/ghostty-org/gh…
English
52
136
2.1K
174.6K
Dexter
Dexter@dexcompiler·
@banteg There's a huge no of devs who basically live for complexity Sad to see
English
1
0
8
1.1K
Dexter
Dexter@dexcompiler·
@burkov Cursor is still largely relevant and will likely continue to be There's a whole swath of devs who don't like working inside the terminal. Might sound unbelievable but its true
English
2
0
2
1.3K
BURKOV
BURKOV@burkov·
Cursor and Perplexity are two products that were only relevant for a short moment in time when the LM tech hadn't matured yet and an army of enthusiasts had nowhere to go. Just like Mistral 7B Instruct was good when it was released but only until the tech industry started to seriously invest in open weight models. It's time to say goodbye to these two. These were a fun three years, but these pet projects aren't relevant anymore.
English
100
22
518
92.7K
Dexter
Dexter@dexcompiler·
@gvanrossum @simonw @rahulj51 It just means "experienced" will be loaded with a different meaning perhaps. Most engineers today have zero understanding of assembly or registers Every higher abstraction presupposes some ignorance unless the individual decides to go through the discomfort of diving deeper
English
0
1
11
1.9K
Rahul Jain
Rahul Jain@rahulj51·
The thing about this is that no one has a clue what human SWEs would be doing instead. The idea that we would all be reviewing code is flawed. Because agents can review code much better. I think our only advantage right now as human SWEs is that we have an almost infinite context window over very long horizons.
Ryan Dahl@rough__sea

This has been said a thousand times before, but allow me to add my own voice: the era of humans writing code is over. Disturbing for those of us who identify as SWEs, but no less true. That's not to say SWEs don't have work to do, but writing syntax directly is not it.

English
83
25
570
102.4K
One Happy Fellow
One Happy Fellow@onehappyfellow·
chatgpt won't write tcp code because of "safety" for fucks sake
One Happy Fellow tweet media
English
122
48
2.3K
144.1K
Dexter
Dexter@dexcompiler·
@banteg Can confirm. Real pain. I was thinking about a weighting and priority system to tag along with the skills but haven't quite gotten around to it yet
English
0
0
0
82
banteg
banteg@banteg·
turned off most skills in codex as they result in agentic degradation. when it reads a skill, it treats it as an authoritative workflow, even if it's just a tiny part of what's needed to provide the answer. adding skills as simple as language cheatsheets (no workflow described, just reference) has reduced my usual run time from 30 mins to 2-3 mins. i just moved everything to docs/ for now.
English
24
6
182
18.8K
Dexter retweeted
Archivara
Archivara@Archivara·
Archivara (powered by 5.2 pro and Opus 4.5) breaks the 2019 benchmark for 5×5 circular (cyclic) matrix multiplication. The best verified construction required rank-8 (8 multiplications); we now provide a practical, fully verified rank-7 algorithm. This realizes a long-known theoretical possibility as an explicit, working method.
Archivara tweet media
English
17
42
371
222.3K
Dexter
Dexter@dexcompiler·
@leerob @Suhail ascii-based architecture diagrams are enough in many cases to communicate the idea simply and powerfully mermaid rendering can be weird sometimes so would be nice to have the ascii option
English
0
0
0
62
Lee Robinson
Lee Robinson@leerob·
@Suhail The plan visual editor is getting even better in the next release! One underrated feature is to ask for mermaid diagrams to visualize architecture.
Lee Robinson tweet media
English
20
4
223
17.1K
Suhail
Suhail@Suhail·
Reading plans from claude code in terminal is super not fun. I don't get how y'all are using it in terminal all the time vs something like Cursor that generates a nice readable markdown file.
English
305
19
766
127K