nekitamo

The Qwen3.8-Flash-Next model is ok for reverse engineering. It can definitely solve simple tasks, but it gets easily confused, and burns tokens like crazy.

I've been sick with stomach flu the past few days, so I put together a small benchmark for all the top AI models where I ask them to reverse engineer a very simple small crackme (https://www.crackmes.one/crackme/6ab4ce5195b976f8f1300b54) using Ghidrasql (https://github.com/0xeb/ghidrasql)

(As an aside: I think representing tools for AI as virtual SQL tables is much better than traditional MCP servers, and that this approach deserves a lot more attention. Props to Elias Bachaalany for thinking out of the box and putting it together).

The results of the benchmark are below. It's sorted by cost, but I added notes about each AI model for myself. The notes try to talk about intelligence, speed, and cost. Intelligence is a subjective measure which I assign to each model by reading its thinking tokens and watching its tool calls. You can definitely tell which models are more intelligent, even on this small benchmarks, by their tool use, false starts, and how easily they fall for the simple misdirection in the crackme.

(gist is more readable on mobile) https://gist.github.com/markoglasgow/8812c19b6fbffed30b2661b...

```

  +----------------------------------+--------------------------------------------------------------------------------+-----------------+-----------------+-----------------+
  | Model / route                    | Experience / outcome                                                           |    Total tokens |   Output tokens | Cost            |
  +----------------------------------+--------------------------------------------------------------------------------+-----------------+-----------------+-----------------+
  | Xiaomi MiMo v2.6 Flash           | Incredibly fast and intelligent; solved.                                       |         225,291 |           3,281 | $0.005 (0.5c)   |
  | GLM 5.3 Flash / OpenRouter       | Least intelligent group but cheapest; solved.                                  |         340,006 |          11,205 | $0.01           |
  | ChatGPT Luna Latest              | Very fast, concise, token-efficient, intelligent, and cheap; solved.           |         401,837 |           5,119 | $0.01           |
  | Qwen 3.8 Flash                   | Initially confused and failed, then eventually solved.                         |        ~418,000 |          16,676 | $0.02           |
  | Qwen 3.8 27B                     | Fast; one wrong answer, then quickly self-corrected.                           |        ~774,784 |          12,658 | $0.02           |
  | DeepSeek Flash / direct API      | Fast, cheap, and accurate; solved.                                             |         369,334 |           9,089 | $0.03           |
  | DeepSeek Flash / OpenRouter      | Fast and intelligent; solved in 76 seconds.                                    |         371,689 |          13,888 | $0.03           |
  | Claude Haiku 5.5                 | Fast and surprisingly intelligent, but overly thorough and token-heavy.        |         662,346 |          26,244 | $0.03           |
  | Thinking Machines Inkling        | Low intelligence; good speed and relatively token-efficient; may loop.         |             N/A |             N/A | $0.03           |
  | Xiaomi MiMo v2.6 Pro             | Overthought the task and wandered into tangents.                               |        ~420,644 |          13,638 | $0.05           |
  | GLM 5.3 / OpenRouter             | Cheap and intelligent but behind DeepSeek; solved in about 2m 20s.             |       ~702,473* |          11,879 | $0.08           |
  | GPT Sol Latest (Sol 6.1)         | Slow, but highly intelligent and extremely token-efficient; solved in ~89s.    |         294,957 |           1,406 | $0.15           |
  | MiniMax M3                       | Least intelligent comparison; many wrong turns and repeated small mistakes.    |      4,546,552* |         116,804 | $0.20           |
  | GPT Terra Latest                 | Very fast, very smart, and token-efficient; solved.                            |         229,888 |           2,342 | $0.20           |
  | Mistral Large 4.0                | Medium speed; small mistakes but reached the solution.                         |         303,687 |           4,041 | $0.21           |
  | Kimi K2.7 Code                   | Less intelligent, slow, expensive; not worth using.                            |         100,740 |           8,938 | $0.22           |
  | Claude Opus 5.5                  | Highest intelligence and token efficiency; pretty fast; cheaper than Sonnet.   |         312,837 |           1,862 | $0.24           |
  | Qwen 3.8 Max 0902                | Fast and intelligent, better than Kimi K3, but token-hungry.                   |       1,305,241 |          26,390 | $0.25           |
  | Kimi K3 / OpenRouter             | Faster and cheaper than direct K3, but still very expensive.                   |         750,064 |           5,922 | ~$0.32          |
  | Meta Muse Spark Open             | Intelligent and medium speed, but among the more expensive runs.               |         872,850 |          10,284 | $0.35           |
  | Gemini 4 Argon                   | Extremely fast and intelligent; strong value for capability.                   |         384,046 |           2,870 | $0.35           |
  | Grok 4.7                         | Very fast and intelligent, but expensive.                                      |         655,648 |           4,217 | $0.46           |
  | Meta Muse Glimmer 30B            | Fast but low intelligence and token-inefficient; cancelled unsolved.           |             N/A |             N/A | ~$0.50 wasted   |
  | Claude Sonnet 5.5                | Smart and fast but very token-inefficient and expensive.                       |      1,243,285* |          20,381 | $0.57           |
  | Kimi K3 / direct API             | Intelligent but not notable; extremely slow and far too expensive.             |         205,710 |           3,215 | $0.72           |
  | ChatGPT Astra Latest             | Medium speed, token-efficient, highly intelligent; cost similar to Fable.      |         278,603 |           1,536 | $0.83           |
  | Claude Fable 5.1                 | OK speed, very expensive, medium token efficiency; solved in ~53s.             |        397,806* |           2,857 | $0.85           |
  | Gemini Flash                     | Refused the reverse-engineering task.                                          |             N/A |             N/A | Not reported    |
  | Thinking Machines Inkling Small  | Got stuck immediately and was terminated.                                      |             N/A |             N/A | Not reported    |
  +----------------------------------+--------------------------------------------------------------------------------+-----------------+-----------------+-----------------+

```

The latest Deepseek-Flash model was the most impressive of the open models. Intelligent, fast, cheap. What more do you want? I think the Xiaomi MiMo - Flash model did well, but it might be a fluke, given how poorly the Xiaomi MiMo Pro model performed (it wandered off on a million side quests).

What's also interesting is how competitive Luna is to the lower end Chinese models. It's quite a bit more intelligent, burns a lot less tokens, is plenty fast and very cheap. After doing this experiment, I don't think it's a foregone conclusion that the American AI labs will be having the bottom of the market eaten away by the Chinese models. It seems they (at least OpenAI) can compete on the bottom end as well.

Finally, the benchmark's results reinforce a few things people have been saying about Sonnet and Terra, namely, they're not competitive for a range of tasks. For easier tasks, Haiku/Luna are plenty capable. For harder tasks, Opus and Sol are so much more intelligent, that they can get them done quicker and cheaper. Maybe there is a place for them in the "Opus designs and Sonnet implements" workflows, but in that case, why not "Opus designs and Deepseek implements?"

PS Code harness used was Claude Code. $20 in openrouter credits was used to access the models. Kimi K3 and Deepseek were also accessed directly via the API and then compared to Openrouter. Kimi K3 specifically was much faster on Openrouter, and the difference in output tokens to get the same results for Deepseek/Kimi makes me think the Openrouter providers are hosting significantly quantized versions of the model. Goes to show that Moonshot.ai's acceptance testing of providers is a useful thing, otherwise your model might look artificially bad when it's hosted as heavily quantized by providers who dont care about quality, just top-line cost and speed.

Could you recommend me a few?

This article is a summary of the author's progress in decompiling CoD: MW2. More info and screenshots can be found on their Twitter feed:

https://x.com/momo5502/status/2100276477065351184?s=46

I think you're on to something. It's really concerning that the people with the money to run these ad campaigns are so removed from the bulk of the populace that they don't even bother targeting them with all the billboards anymore, just each other.

Otoh maybe I'm looking at it wrong, and maybe billboard advertising has gotten so ineffective that it's cheapened to the point that you can tuck it into your campaign as an afterthought. If you get one solid enterprise lead from it, it pays for the whole year. And besides it has branding value.

Either way, definitely feels a bit dystopian to me. How alienating must it feel for the service worker commuting to his dead end job where even the billboards don't make sense anymore?

This tracks, in my experience the 27B is better at coding and instruction following. I'm shocked at how much of a difference the dense models vs MoE makes.

But it's a moot point, because for local inference on consumer hardware, the MoE is so much faster.

Yes, Wine was (EDIT: partially) compatible with the Roblox anti-cheat, and there was a period in time where Roblox with its anti-cheat ran on Wine. Eventually, however, a decision was made to block Wine.

EDIT It looks like the Roblox anti-cheat devs themselves explained why Wine was blocked:

https://devforum.roblox.com/t/roblox-on-linux-everything-sum...

I personally liked their Code virtualizer. And it's under active development. Sorry to hear about Theia, I haven't run into it yet.

This isn't true. Roblox never had and does not have a kernel anti-cheat.

I don't think that's necessarily true. Certainly the video streaming is expensive, but other things like workers and KV store are quite cheap and performant and powerful _if you structure your access patterns accordingly_ .

Cloudflare pricing introduces an additional dimension for when you're architecting software on top of them, but if you do it correctly, your product has the potential to be faster, cheaper, and easier to run than traditional solutions running on multiple geographically distributed VMs. You just can't approach them as if they're just another VM instance provider and expect a similar experience and pricing. What they do and price for is fundamentally different from that.

They are still around, under the name Denuvo :)

You can protect human rights without stifling progress. It's not a "pick one of the above" situation.

The EU can and should reform many parts of its sclerotic laws and bureaucracies. Whether it can do so before it becomes a subservient puppet state which serves as a battleground for competing powers remains to be seen.

While AI isn't particularly good with obfuscated code, it is true that code obfuscation in most old games can be removed relatively easily. The attack and defense state-of-the-art has moved on from those days.

Given that 42% of Android devices are unpatched as of now [1] it's an interesting decision on their part to release their research and make them all vulnerable

[1] https://gs.statcounter.com/android-version-market-share [2] https://www.cybersecurity-insiders.com/survey-reveals-over-1...

I wonder if this is related to the pretty serious security incident about Github which got published today:

https://x.com/sagitz_/status/2049153195243372569

With malicious HTTP headers, any user could access any repo on Github.com, or on the Enterprise Github instance they might have access to. It's even worse than that because it's remote code execution on the Github server.

It seems like Github has been a mess since the Microsoft acquisition. Definitely feels like another multi billion dollar screwup in the making, like Skype or Nokia were.

Hopefully the incidents in the last few weeks are a wakeup call, and they start getting their shit together.

That's interesting, Kimi K2.5 used through KimiCode was comparable to Sonnet in my tests, and is an excellent alternative to Anthropic models

That being said, I noticed that Kimi being served through Openrouter providers was trash. Whatever they do on the backend to optimize for throughput really compromised the intelligence of the model. You have to work with Kimi directly if you want the best results, and that's also probably why they released a test suite to verify the intelligence of their new models.

Ideally you would want someone like Charles De Gaulle.

The job is inherently difficult, and I think a big problem is institutional decay/drift leading to a bad pipeline of leaders, which is why you have so many poor, weak, and ineffectual leaders serving back to back. The UK is a prime example of this, but I think all of Western societies are struggling with this to one degree or another.

Indeed hindsight is 20/20, and I won't pretend to have all the answers. I just personally think we have a particularly rotten batch of leaders which can't only be explained by the leaders themselves, but also by the institutions and policies which spawned those leaders.

I agree the US has many problems, and I really don't want to make this a EU vs USA thread. I also wouldn't say the US is "successful", whereas the EU is not. I just think the EU has amazing potential and isn't living up to it.

Also I think any success the US does enjoy over the EU is in spite of the things you mentioned, and a large part of that is the US simply has a much larger economy, much more money, and much deeper and well developed capital markets. Which just goes to show how much more the EU could aspire to, being a much larger bloc of countries with a larger population and all.

I'm sorry, I didn't mean to engage in tropes, I'll be more specific.

Emmanuel Macron is so unpopular that he has to forestall elections until 2029 to remain in power. In the meantime, his unpopularity means France cannot form a government or pass a budget, while the political center erodes under his leadership, giving way to the far left and far right.

Angela Merkel presided over a disastrous energy policy (outsource the coal mines to Poland, close all nuclear reactors, rely on cheap energy from Russia) which made German industry, and Europe by extension, precariously dependent on outside partners which are proving to be very unreliable. This has resulted in reduced economic performance, increased consumer costs, leading to popular discontent. This coupled with a poorly thought out immigration policy are hallmarks of her time in power, the fallout of which Germany is still dealing with today.

Ursula von Der Leyen is a direct descendant of Merkel, but with much less to show. She was complicit in all of Merkel's poor policies, and has not been able to address any of their negative consequences effectively. She has failed to rearm Europe, she has failed to revive economic growth (indeed, just the opposite, embracing at times a de-growth agenda which might on paper be noble, it incompatible with our current economic systems), she has done nothing to reassert Europe's sovereignty in matters of defense and energy, she has presided over the worst excesses of the European Council, which counter-productively rob individual countries of their sovereignty through a combination of bad lawmaking and policy (see the Draghi report), and poor executive decisions (see EU forcing Poland and Romania to buy $2 billion of vaccines they don't need and didn't ask for last week).

I won't even get into Donald Tusk, Viktor Orban, Karol Nawrocki, Hollande, Sarkozy, and all the pre-2020 Italian prime ministers (special shout out to Berlusconi lol).

Finally, I respect that it might be bad to engage in tropes, but I think it's also frame any criticism as playing to the far-right. Indeed a big problem in Europe is centrist politicians have suffocated any criticism by labeling it as "far-right". Over time, as their incompetence leads to more criticism, they label more people as far-right. This has had the reactionary effect of pushing otherwise normal centrist people into the far-right camp, which explains the rise of Le Penn and AfD, to the point that about 25% of voters in France/Germany are unfortunately voting for these far-right options.

Any healthy society must allow for debate and criticism, without labeling everyone who disagrees as extremists.

As an American living in Europe, I don't think the well-balanced European way of life is the cause of Europe "falling behind". Instead I think it's a combination of the following intertwined factors: bad policies, a stunningly incompetent array of bad leaders, and bad deployment of capital (by both private investors and the state).

Agreed otherwise, the essay is great.

One of their journalists also doxxed Naomi Wu, intruding on her personal life, making her lose her income, and possibly getting her in trouble with Chinese authorities: https://x.com/RealSexyCyborg/status/1209815150376574976

The journalist themselves is a real piece of work: https://thehill.com/homenews/media/463503-sarah-jeong-out-at...

Kinda goes to show you the kind of people who write these stories. Ethics haven't been on their mind for a long time, and them preaching to anyone about ethics is rank hypocrisy.

100% this. Ant was bad in many ways, but at least it was lightning fast. Gradle is just tragic.

I'd like to mention the web framework I'm using these days, Jooby:

https://jooby.io/

I've found it quite satisfying compared to the other "new" ones.

As for the original topic, I just want to echo what others have said, and say that I am happiest in Java when writing it as if it were Golang code. That an the first-class runtime and performance and deep ecosystem make it a great choice in 2026.

I've participated in some corporate shit-shows in my day, but man I don't think I've ever seen one burn cash this fast.

Another thought: they say the software you ship reflects your org chart ("you ship your org chart"). Given how far Meta has slipped in the last year in the AI race, their org-wide dysfunction is starting to seriously harm them, from Financials to execution to talent. They need to get their act together, starting from the top.

I'm not a fan of Meta, but I'm a big fan of Llama. It was the first notable open weights model, and paved the way for all the others. Just for that I want to say: I'm rooting for you guys. Hope an amazing Llama 5 release comes after all this pain and churn.

Which comparable US dedicated server providers do you prefer?

In my tests, Qwen3.5-35B-A3B is better, there is no comparison. Better tool calling and reasoning than Qwen3-Coder-Next for Html/Js coding tasks of medium size. Beware the quants and llama.cpp settings, they matter a lot and you have to try out a bunch of different quants to find one with acceptable settings, depending on your hardware.

Thank you. The difference was quite noticeable today.

Because then they get some form of control over Anthropic. Solely through the act of using it, they claim some form of ownership over it.