Agree, many people think code reviews are there to catch "mistakes" (I used to think so too) but I think in most cases, they don't work that way. Bugs are still shipped and, at least in my experience, most reviewers often just quickly skim the changes without paying attention to the details. And those code reviews only give false confidence that a change is OK and must be error-free and good to go just because it was given LGTM by a senior dev (who barely looked at it), with all the consequences of such false confidence. It feels like in 95% of cases review is not even necessary and is just a corporate ritual at this point. It makes sense when you have a fresh contributor and you don't have confidence in them yet (i.e. code review as an onboarding process) or when a certain feature is non-trivial and you aren't sure everyone will be on the same page about the architecture (to prevent future conflicts/disagreements/rewrites of sorts), but other than that, proper QA processes are more important and more effective than CR at catching mistakes (tests, canaries, manual testing too etc.). Our company used to have a strong CR culture but now for a few years, there've been talks that CR is a bottleneck which doesn't really solve the problems it's supposed to solve.
My current pipeline with AI-generated code (and my own code too) is that I run AI reviewers several times on a single PR (the way LLMs work, due to attention optimizations, quantization issues, token sampling issues, harness issues -- AI agents often overlook some things on the first attempt), then I assemble the final report and we quickly decide together what needs to be addressed (critical stuff) and what's okay as is. I quickly skim the change set myself too, to see what the shape of the tests is, and what code areas were touched, just to be informed, and that's about it. I don't feel exhausted personally, because it pays off when there's a working feature at the end of the day
Not sure if you need this Artex tool at all? I was able to find 15 vulnerabilities including an RCE in 2 casual nights by just saying "go hack it" to a pretty good model with weak guardrails. No tooling needed except for a Linux environment
Say, today I found a service had image uploads. I told it, go check that stuff. One of its hypotheses was a decompression bomb. 10 minutes later the service gives 503's and is stuck in CrashLoopBackoff. It has all the knowledge and tools to craft a bad PNG.
Or yesterday there was this article about Lovable users not securing their Supabase DB's properly. I told it to find a couple Lovable sites and probe them for exposed tables (no RLS). At first it refused to do it (the guardrails activated) but then it agreed to list only 1 row from each table with PII redacted as a compromise for my PoC. I then just opened a new session and asked it to remove PII redaction and return all rows. It complied becase it didn't pay attention to what the script was about. One of the Lovable sites exposed user passwords in plaintext.
Basically if my goal is "hack something successfully today" (by success I mean DoS, data breach or RCE), the success rate so far has been 100% (at least one site is compromised). It's scary how easy it is nowadays
People say GLM 5.3 is not much smarter than GLM 5.3 Flash (just +4 points on Artificial Analysis), and GLM 5.3 Flash fits on 4x RTX 6000 Pro for 4 times less money, with the same 600 tok/sec aggregate decode speed on vLLM (6 mln tokens KV cache). And I tell you, 600 tok/sec is not much for an organization, it's barely usable for 8 people. That's why we plan to buy more RTX 6000 Pros. And this setup requires $400,000 to run GLM 5.3 for 16 people, it's absolutely uneconomical.
5.3 is certainly “better” than its flash variant, in my experience. I’ve moved on from both but eagerly await the next generation of each.
Better, in this context, is more capable with open ended work, planning, longer horizon “reasoning”. Flash is an absolute chad with a well defined task and accompanying guidance/guardrails/conventions provided. But not even close when it needs to deviate, runs into a “blocker” and has to reason its way out.
Personally I find GLM-5.3 to be smart enough to be my daily driver, writing code, doing investigations, etc., including multi-codebase changes and finding the relevant documentation/supporting code for itself so I don't have to explain everything. It's fast (depending on your inference provider), cheap, and smart enough.
I find the full-size 5.3 is smart enough to avoid all the small-model problems, while GLM-5.3-Flash definitely starts showing its limits any time the task at hand is poorly defined. 5.3-Flash can often manage even complicated tasks, but the error rate (in terms of bad decisions, misunderstandings, etc) starts creeping up and the time wasted (or the risk thereof) gets high enough that it's worth just paying for the bigger model.
I do wish the bigger model could use vision though- that's currently limited to the flash model (presumably because it came out later, when they had figured out their vision training pipeline. Hopefully later releases will include it.)
Dunno, maybe I'm good at prompting it, but I haven't had many issues with GLM-5.3 Flash. After switching from the Claude models, I barely noticed any difference for most tasks. In fact, I have to recheck the settings from time to time to make sure I'm not running anything Anthropic when I see it working with sensitive data :)
The only annoying issue for me is that GLM-5.3 Flash is often eager to go into action and start modifying code when I've just asked a question and it misunderstood it as a call to action.
People say GLM 5.3 is not much smarter than GLM 5.3 Flash (just +4 points on Artificial Analysis)
Not the only case where AA's numbers have no correlation to actual user experience. The full GLM 5.3 model is vastly smarter than the Flash version, and vastly slower.
The setup in the article reports 600 tok/sec aggregate decode for GLM 5.3. I get the same throughput on 4x RTX 6000 Pro's connected vi PCIe for GLM 5.3 Flash, which is just 4 points below GLM 5.3 on Artificial Analysis. I'm just saying the setup in the article looks overpriced for what it can actually do
>the real risk is the default world tech stack shifts to open models on Chinese chips.
As someone who isn't based in the US/EU and whose job is to deploy LLMs on premises, I can't wait for the moment when China starts selling its GPUs globally en masse with a proper CUDA equivalent. It's going to be cheaper in the long term and without all the export control drama.
Which is a good thing. I think it's insane to see so many comments along the lines of, we have to keep China from becoming the world leader in GPUs and AI tech instead of the US. Really? Why? Look at the evil the US has used its AI for. It's time for a more ethical nation to take the lead, and better for the world.
>Due to improper synchronization when executing concurrent threads sharing resources, an attacker can trigger heap corruption.
>As of the time of this blog’s publication, no exploit is available and there are no reports of exploitation in the wild.
Since when does "memory corruption" = "RCE"? Last week I found and reproduced a memory corruption (a use-after-free) in PHP's runtime after running an on-premise agent for just 30 minutes. There's nothing about it on the internet. Should I write a blog post about a scary new vulnerability? :)
I was prepared to disagree with you, but I agree: I'd say there's a strong norm of not calling things RCE unless you can actually demonstrate RCE, or at least have a clean pathway to demonstrating it (e.g. a full write-what-where).
It's perfectly sufficient to have a CVE for just memory corruption, so the "RCE" framing here feels overstated. Particularly when an LLM should easily prove RCE out if it is in fact present.
> (...) public research has demonstrated a reliable exploit against a default-configured server with the mod_http2 module enabled. Successful exploitation allows an attacker to execute arbitrary commands, gain full control of the server process, and access or modify sensitive information.
I don't know the backstory here, but there's a long and colorful history of vendors and other parties guessing that some contrived memory corruption bug isn't exploitable and then being proven wrong, so the industry now more or less defaults to the assumption that it's exploitable.
Let me remind you that in response to the PHP bug, one of the front-page HN posts was a guy saying it's probably not exploitable and another offering Bitcoin to anyone who pops their server... with predictable results.
People can be pretty creative with exploitation primitives, and LLMs now replicate that at a scale. And yep, if you found a use-after-free that can be triggered remotely in PHP under some reasonable assumptions, you should probably treat it as a security bug.
If you can corrupt memory in a controllable way, you can get a computer to do just about anything that process has rights to. Possibly even the kernel, if you can get at some shared memory.
Presumably one of Anthropic, Google, or Apache possesses a working exploit, because the official Apache Foundation report on this CVE claims RCE is possible.
Last week I deployed GLM-5.3-Flash on our on-premise servers, ran it against all repos, and in 6 hours it found 13 vulnerabilities across our products/sites, including 1 full account takeover and an RCE on our site which involved a pretty sophisticated chain of __destruct gadgets (all vulns verified to work by a human).
As an experiment, I also pointed it at the PHP runtime's source code, and it was able to construct PHP code that triggered a use-after-free (or a segfault, if you prefer DoS). I failed to find the vulnerability in the list of CVEs, so maybe it's a 0day, I don't know.
Sure, it's not a frontier model, but it can find many issues that were overlooked by all the commercial security scanners we run regularly as well. The nice thing about this setup is that we can run it 24/7 probing our code/infra, constructing exploits, etc. without having to care about token usage/bills. Plus, I think sending all that sensitive information about your internal infrastructure and the vulns found to Anthropic is a vulnerability in itself.
It takes 116 seconds to process 43 input tokens and produce an answer to a very simple request (translation). That's pretty unusable, to be honest. Is overthinking a trait of DeepSeek, or it's a low quantization issue?
For comparison, the whole thing takes 1.8 sec on my RTX 6000 rig with GLM5.3-Flash, and 12.5 seconds on my RTX 5090 rig with Qwen 3.6 27b (it overthinks too, unlike GLM).
I can't imagine waiting 2 minutes for each small turn. It's like... 1 coding task per night? Sounds like, as an engineer, I can accomplish more in that time if I did everything myself manually without any AI.
That's not the speed of the inference, but the fact that the default parameters are to max thinking. You can even disable thinking for many queries, or lower it to a smaller value.
Yes, Deepseek consumes a higher number of tokens per turn that comparable models (the latest Muse for example). This is just my experience, not a scientific study.
I'm currently evaluating GLM and oh boy, lots of correctness issues on large contexts if you run it on vLLM (bugs in the engine itself). I wonder what their config is.
The average age of the explored models is ~1.5 years. They discuss PHP 8.3, which was released in 2023 and has not been actively supported since 2025. Agentic coding isn't used.
I'm not sure what the point is of publishing results on the mid-2025 state of the art in October 2026.
My current pipeline with AI-generated code (and my own code too) is that I run AI reviewers several times on a single PR (the way LLMs work, due to attention optimizations, quantization issues, token sampling issues, harness issues -- AI agents often overlook some things on the first attempt), then I assemble the final report and we quickly decide together what needs to be addressed (critical stuff) and what's okay as is. I quickly skim the change set myself too, to see what the shape of the tests is, and what code areas were touched, just to be informed, and that's about it. I don't feel exhausted personally, because it pays off when there's a working feature at the end of the day
reply