(Which I still think is bonkers, FWIW.)
Opus 4.6 is an especially interesting comparison point, I think, because it was a step change; IMO it's when LLMs became serviceable for coding. Yeah, pre-4.6 models did output code, and that code often superficially worked; and bringing it up to production standards still generally meant rewriting it entirely. Opus 4.6 is when that changed.
From my early tests, it's looking like the public benchmarks are not misleading, and Qwen 3.8 somehow got there too, by and large. I've got a few personal tests. One is a mid-complexity one-shot, purposefully underspecified. Beyond a few minor bugs that it could easily fix once pointed out, Qwen 3.8 largely aced it. There are a bunch of things I'd improve, but that was true of Opus 4.6's output too, and by and large the code is clean and well structured. Also worth noting that I'm running Qwen 3.8 fairly aggressively quantized to fit in VRAM; I'd expect tighter results still from the full weights.
Another test I ran is a variant of a common puzzle with an additional structural constraint that makes the usual solution inapplicable, so the model has to actively turn away from the well-known solution and construct a new one that takes the constraint into account. I've never seen a home model pass that test. Kimi K3 passes it, GLM 5.2 passes it (painstakingly). Qwen 3.8 struggles a lot... but does arrive at a correct solution. First time I see a home model do so.
I haven't yet tested it on long multi-turn scenarios. In my experience, that's where pocket models are weakest against heavyweight ones, especially when quantized.
That aside, it does seem like Qwen 3.8 can, in fact, trade blows with Opus 4.6. I don't know yet if it could replace it, and my money would be on no, but I may well be wrong about that considering how weirdly capable it is.
Interestingly, Qwen 3.8's MTP layer is uncannily accurate too. It still gave me good results up to 6 to 8 predicted tokens, which boosts its speed so much it's competitive with Qwen 3.6 MoE. So that's another bizarrely impressive thing about it.
And given all of the above, I do think that the trillions of dollars invested into OpenAI and Anthropic are becoming harder and harder to justify."},"story_title":{"matchLevel":"none","matchedWords":[],"value":"Qwen 3.8 27B"},"story_url":{"matchLevel":"none","matchedWords":[],"value":"https://huggingface.co/Qwen/Qwen3.8-27B-FP8"}},"_tags":["comment","author_Balinares","story_49299605"],"author":"Balinares","comment_text":"That's a reasonable point, so let me qualify: at a glance, it seems Qwen 3.8 27B can trade blows with Opus 4.6 on coding tasks, where trading blows doesn't necessarily mean it's a clear winner or even an equal, but does mean it'll at least hold its own and land a punch or two.
(Which I still think is bonkers, FWIW.)
Opus 4.6 is an especially interesting comparison point, I think, because it was a step change; IMO it's when LLMs became serviceable for coding. Yeah, pre-4.6 models did output code, and that code often superficially worked; and bringing it up to production standards still generally meant rewriting it entirely. Opus 4.6 is when that changed.
From my early tests, it's looking like the public benchmarks are not misleading, and Qwen 3.8 somehow got there too, by and large. I've got a few personal tests. One is a mid-complexity one-shot, purposefully underspecified. Beyond a few minor bugs that it could easily fix once pointed out, Qwen 3.8 largely aced it. There are a bunch of things I'd improve, but that was true of Opus 4.6's output too, and by and large the code is clean and well structured. Also worth noting that I'm running Qwen 3.8 fairly aggressively quantized to fit in VRAM; I'd expect tighter results still from the full weights.
Another test I ran is a variant of a common puzzle with an additional structural constraint that makes the usual solution inapplicable, so the model has to actively turn away from the well-known solution and construct a new one that takes the constraint into account. I've never seen a home model pass that test. Kimi K3 passes it, GLM 5.2 passes it (painstakingly). Qwen 3.8 struggles a lot... but does arrive at a correct solution. First time I see a home model do so.
I haven't yet tested it on long multi-turn scenarios. In my experience, that's where pocket models are weakest against heavyweight ones, especially when quantized.
That aside, it does seem like Qwen 3.8 can, in fact, trade blows with Opus 4.6. I don't know yet if it could replace it, and my money would be on no, but I may well be wrong about that considering how weirdly capable it is.
Interestingly, Qwen 3.8's MTP layer is uncannily accurate too. It still gave me good results up to 6 to 8 predicted tokens, which boosts its speed so much it's competitive with Qwen 3.6 MoE. So that's another bizarrely impressive thing about it.
And given all of the above, I do think that the trillions of dollars invested into OpenAI and Anthropic are becoming harder and harder to justify.","created_at":"2026-08-16T11:35:42Z","created_at_i":1786880142,"objectID":"49319073","parent_id":49302610,"story_id":49299605,"story_title":"Qwen 3.8 27B","story_url":"https://huggingface.co/Qwen/Qwen3.8-27B-FP8","updated_at":"2026-08-16T11:47:58Z"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"gymbeaux"},"comment_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["llm","benchmarks","are","misleading"],"value":"LLMs have taken much of the enjoyment out of coding for me, but I don\u2019t think it has to be that way. I hope that as an industry we settle on LLMs being more like tools than human assistants. I think most of the engineers at my company are using LLMs as human assistants- most of them have agentic workflows set up and have premium subscriptions to Claude AND Gemini AND ChatGPT. Many have local LLMs running on their company MacBook Pros, but they can never manage to describe to me in plain English what those local LLMs are doing. I would compare local LLMs to Raspberry Pi clusters. They\u2019re neat, and technically they can do stuff but they are incredibly impractical. I want so desperately to have the power of Claude Opus running locally, but we are incredibly, ridiculously, extremely far off (benchmarks are often misleading and nobody likes to talk about the paltry tokens/s they\u2019re getting on their home rig).
So the problem for me is that I\u2019ve noticed a trend with my coworkers- they usually don\u2019t have a formal CS background (e.g., Comp Sci degree), their resumes are unimpressive, and they have personal websites where anyone can download a copy of their resume. I haven\u2019t seen this at any previous job. It\u2019s just weird, man. Anyway, so my problem is when they are purported to be the best of the best at this company, and they\u2019re chugging the LLM kool-aid, I\u2019m struggling to figure out whether they\u2019re really ahead of the curve and getting the drop on the rest of us, or they\u2019re doing basically no work and waving through each other\u2019s PRs with little to no oversight on what the LLMs are writing. The bush league nature of many of the bugs I\u2019m seeing going all the way to prod makes me think it\u2019s the latter.
Meanwhile, these guys and their teams are completing an absurd number of story points each sprint. My team is dead last (and coincidentally we\u2019re all pretty anti-AI for writing code).
I\u2019m waiting for the other shoe to drop. I\u2019m waiting for the C suite to call our project a failure and lay everyone off. The product generally works, though I think if the devs were truly rockstars, things would be much smoother. It seems like LLMs are letting middle-of-the-road or mediocre devs appear to be rockstars purely on output speed. Nobody in this org cares if the output has 4x the bugs it should have, just that it was completed quickly.
The main issue I think is that nobody- including the rockstar tech leads- knows how the system works. They don\u2019t know the code. There\u2019s a bug and it takes forever to track down. Don\u2019t ask about unit tests.
I think in a perfect world, these LLMs are treated like another tool rather than another person- meaning you and I still write a large portion of the code, and we read and understand every line- especially the lines generated by LLMs. It\u2019s still a huge productivity multiplier for me to not have to remember how to do mundane things like write out arrays into CSV files- before I\u2019d find a close implementation on StackOverflow and tweak to my needs. Now I can just describe what I want the CSV to look like and Claude Code generally gets it right on the first try."},"story_title":{"matchLevel":"none","matchedWords":[],"value":"What's gonna happen to software engineers?"},"story_url":{"matchLevel":"none","matchedWords":[],"value":"https://yakko.dev/blog/whats-gonna-happen-to-software-developers"}},"_tags":["comment","author_gymbeaux","story_48364284"],"author":"gymbeaux","comment_text":"LLMs have taken much of the enjoyment out of coding for me, but I don\u2019t think it has to be that way. I hope that as an industry we settle on LLMs being more like tools than human assistants. I think most of the engineers at my company are using LLMs as human assistants- most of them have agentic workflows set up and have premium subscriptions to Claude AND Gemini AND ChatGPT. Many have local LLMs running on their company MacBook Pros, but they can never manage to describe to me in plain English what those local LLMs are doing. I would compare local LLMs to Raspberry Pi clusters. They\u2019re neat, and technically they can do stuff but they are incredibly impractical. I want so desperately to have the power of Claude Opus running locally, but we are incredibly, ridiculously, extremely far off (benchmarks are often misleading and nobody likes to talk about the paltry tokens/s they\u2019re getting on their home rig).
So the problem for me is that I\u2019ve noticed a trend with my coworkers- they usually don\u2019t have a formal CS background (e.g., Comp Sci degree), their resumes are unimpressive, and they have personal websites where anyone can download a copy of their resume. I haven\u2019t seen this at any previous job. It\u2019s just weird, man. Anyway, so my problem is when they are purported to be the best of the best at this company, and they\u2019re chugging the LLM kool-aid, I\u2019m struggling to figure out whether they\u2019re really ahead of the curve and getting the drop on the rest of us, or they\u2019re doing basically no work and waving through each other\u2019s PRs with little to no oversight on what the LLMs are writing. The bush league nature of many of the bugs I\u2019m seeing going all the way to prod makes me think it\u2019s the latter.
Meanwhile, these guys and their teams are completing an absurd number of story points each sprint. My team is dead last (and coincidentally we\u2019re all pretty anti-AI for writing code).
I\u2019m waiting for the other shoe to drop. I\u2019m waiting for the C suite to call our project a failure and lay everyone off. The product generally works, though I think if the devs were truly rockstars, things would be much smoother. It seems like LLMs are letting middle-of-the-road or mediocre devs appear to be rockstars purely on output speed. Nobody in this org cares if the output has 4x the bugs it should have, just that it was completed quickly.
The main issue I think is that nobody- including the rockstar tech leads- knows how the system works. They don\u2019t know the code. There\u2019s a bug and it takes forever to track down. Don\u2019t ask about unit tests.
I think in a perfect world, these LLMs are treated like another tool rather than another person- meaning you and I still write a large portion of the code, and we read and understand every line- especially the lines generated by LLMs. It\u2019s still a huge productivity multiplier for me to not have to remember how to do mundane things like write out arrays into CSV files- before I\u2019d find a close implementation on StackOverflow and tweak to my needs. Now I can just describe what I want the CSV to look like and Claude Code generally gets it right on the first try.","created_at":"2026-06-02T02:28:29Z","created_at_i":1780367309,"objectID":"48365239","parent_id":48364284,"story_id":48364284,"story_title":"What's gonna happen to software engineers?","story_url":"https://yakko.dev/blog/whats-gonna-happen-to-software-developers","updated_at":"2026-06-04T09:31:30Z"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"maeil"},"comment_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["llm","benchmarks","are","misleading"],"value":"I'm building a product that requires complex LLM flows and out of OpenAI's "cheap" tier models, the old versions of Turbo-3.5 are far better than the last versions of it and 4o-mini. I have a number of tasks that the former consistently succeed at and the latter consistently fail at regardless of prompting.
Leaderboards and benchmarks are very misleading as OpenAI is optimizing for them, like in the past when certain CPU manufacturers would optimize for synthetic benchmarks.
Fwif these aren't chat usecases, for which the newer models may well be better."},"story_title":{"matchLevel":"none","matchedWords":[],"value":"Structured Outputs in the API"},"story_url":{"matchLevel":"none","matchedWords":[],"value":"https://openai.com/index/introducing-structured-outputs-in-the-api/"}},"_tags":["comment","author_maeil","story_41173223"],"author":"maeil","comment_text":"I'm building a product that requires complex LLM flows and out of OpenAI's "cheap" tier models, the old versions of Turbo-3.5 are far better than the last versions of it and 4o-mini. I have a number of tasks that the former consistently succeed at and the latter consistently fail at regardless of prompting.
Leaderboards and benchmarks are very misleading as OpenAI is optimizing for them, like in the past when certain CPU manufacturers would optimize for synthetic benchmarks.
Fwif these aren't chat usecases, for which the newer models may well be better.","created_at":"2024-08-07T04:41:59Z","created_at_i":1723005719,"objectID":"41178291","parent_id":41177796,"story_id":41173223,"story_title":"Structured Outputs in the API","story_url":"https://openai.com/index/introducing-structured-outputs-in-the-api/","updated_at":"2024-09-20T17:37:06Z"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"fancyfredbot"},"comment_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["llm","benchmarks","are","misleading"],"value":"If you game the benchmark then you always get found out by your users. Yet the practice remains common in hardware. Outright lies are uncommon but misleading and cherry picked numbers are pretty much standard practice.
The fact that misleading benchmarks don't even drive profit at Meta didn't seem to stop them doing the same thing, but perhaps this isn't very surprising. I imagine internal incentives are very similar.
Unlike the hardware companies though, gaming the benchmark in LLMs seems to involve making the actual performance worse, so perhaps there is more hope that the practice will fade away in this market."},"story_title":{"matchLevel":"none","matchedWords":[],"value":"Llama 4 Smells Bad"},"story_url":{"matchLevel":"none","matchedWords":[],"value":"https://fastml.com/llama-4-smells-bad/"}},"_tags":["comment","author_fancyfredbot","story_43779915"],"author":"fancyfredbot","comment_text":"If you game the benchmark then you always get found out by your users. Yet the practice remains common in hardware. Outright lies are uncommon but misleading and cherry picked numbers are pretty much standard practice.
The fact that misleading benchmarks don't even drive profit at Meta didn't seem to stop them doing the same thing, but perhaps this isn't very surprising. I imagine internal incentives are very similar.
Unlike the hardware companies though, gaming the benchmark in LLMs seems to involve making the actual performance worse, so perhaps there is more hope that the practice will fade away in this market.","created_at":"2025-04-24T07:41:00Z","created_at_i":1745480460,"objectID":"43780162","parent_id":43779915,"story_id":43779915,"story_title":"Llama 4 Smells Bad","story_url":"https://fastml.com/llama-4-smells-bad/","updated_at":"2025-04-24T08:08:27Z"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"kisjovan"},"story_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["llm","benchmarks","are","misleading"],"value":"Hey everyone,
Jovan from UkisAI here! Today, we are introducing Swift, a family of efficient reasoning LLMs based on Qwen, trained by penalizing tokens related to pathological overthinking patterns and restoring accuracy via RL and OPD.
After amazing feedback and 350k+ downloads in 13 days on our Swift Qwen 3.8 27B we are releasing the entire model family as well as the highly requested GSQ-RCO quants for 27B and Flash-Next.
This release includes:\nSwift1.5 27B, an improved version of our last model, with even lower token usage, fixed bugs and better agentic performance, with -58.5% thinking tokens while scoring 0.35% higher while outperfoming base on Terminal Bench 2.1 by not falling into "overthinking error" loops\nSwift Flash Next, with 63.4% fewer thinking tokens and a 1.8x speed up scoring -0.2% vs base on xhigh\nSwift Bonsai 2, with 39.8% fewer thinking tokens while scoring 0.19% higher (although we'd still like to note it as experimental)
Our benchmarks are ran x5 on Base and Swift, averaging across five seeds and various domains, including General (GPQA, AIME26), Coding (LiveCodeBench), Vision (ERQA), Agentic (Terminal Bench 2.1). \nOne note is that the Terminal Bench 2.1 scores are misleadingly low at first glance. It is not a bug, but a simple matter of the Swift models not falling into overthinking loops and failing the task, rather pursuing it until the end, leading to higher average token usage. The token reduction still falls in the -38.7% range when compared apples-to-apples.
We are including a Free Research API and HuggingFace Spaces to give the models a spin before downloading or if you don't have enough compute to run them right now! You can find both on the model cards.
We have also made GGUF, NVFP4, MLX and W4A16 quants for relevant model versions.
More details on our training approach and community feedback can be seen here: https://www.reddit.com/r/LocalLLaMA/comments/1wg7dd5/ukisai_swiftqwen3827b_583_thinking_x195_speed/
All of the various quantization and model versions are available in their respective collections:\nSwift1.5 27B: https://huggingface.co/collections/ukisai/swift-15-27b\nSwift Flash Next: https://huggingface.co/collections/ukisai/swift-flash-next\nSwift Bonsai 2: https://huggingface.co/collections/ukisai/swift-bonsai-2
We would greatly appreciate your feedback via independent evaluations. As per last release, we operate on a candy-shop basis, trying to fulfill as many Swift model requests and quants as possible, so please do share your needs in the comments!"},"title":{"matchLevel":"none","matchedWords":[],"value":"UkisAI Swift Series / 27B, Flash Next and Bonsai 2 /-63.4% thinking, x1.95 speed"}},"_tags":["story","author_kisjovan","story_49833102","ask_hn"],"author":"kisjovan","children":[49833439],"created_at":"2026-09-24T16:33:21Z","created_at_i":1790267601,"num_comments":1,"objectID":"49833102","points":1,"story_id":49833102,"story_text":"Hey everyone,
Jovan from UkisAI here! Today, we are introducing Swift, a family of efficient reasoning LLMs based on Qwen, trained by penalizing tokens related to pathological overthinking patterns and restoring accuracy via RL and OPD.
After amazing feedback and 350k+ downloads in 13 days on our Swift Qwen 3.8 27B we are releasing the entire model family as well as the highly requested GSQ-RCO quants for 27B and Flash-Next.
This release includes:\nSwift1.5 27B, an improved version of our last model, with even lower token usage, fixed bugs and better agentic performance, with -58.5% thinking tokens while scoring 0.35% higher while outperfoming base on Terminal Bench 2.1 by not falling into "overthinking error" loops\nSwift Flash Next, with 63.4% fewer thinking tokens and a 1.8x speed up scoring -0.2% vs base on xhigh\nSwift Bonsai 2, with 39.8% fewer thinking tokens while scoring 0.19% higher (although we'd still like to note it as experimental)
Our benchmarks are ran x5 on Base and Swift, averaging across five seeds and various domains, including General (GPQA, AIME26), Coding (LiveCodeBench), Vision (ERQA), Agentic (Terminal Bench 2.1). \nOne note is that the Terminal Bench 2.1 scores are misleadingly low at first glance. It is not a bug, but a simple matter of the Swift models not falling into overthinking loops and failing the task, rather pursuing it until the end, leading to higher average token usage. The token reduction still falls in the -38.7% range when compared apples-to-apples.
We are including a Free Research API and HuggingFace Spaces to give the models a spin before downloading or if you don't have enough compute to run them right now! You can find both on the model cards.
We have also made GGUF, NVFP4, MLX and W4A16 quants for relevant model versions.
More details on our training approach and community feedback can be seen here: https://www.reddit.com/r/LocalLLaMA/comments/1wg7dd5/ukisai_swiftqwen3827b_583_thinking_x195_speed/
All of the various quantization and model versions are available in their respective collections:\nSwift1.5 27B: https://huggingface.co/collections/ukisai/swift-15-27b\nSwift Flash Next: https://huggingface.co/collections/ukisai/swift-flash-next\nSwift Bonsai 2: https://huggingface.co/collections/ukisai/swift-bonsai-2
We would greatly appreciate your feedback via independent evaluations. As per last release, we operate on a candy-shop basis, trying to fulfill as many Swift model requests and quants as possible, so please do share your needs in the comments!","title":"UkisAI Swift Series / 27B, Flash Next and Bonsai 2 /-63.4% thinking, x1.95 speed","updated_at":"2026-09-24T16:53:23Z"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"intenex"},"comment_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["llm","benchmarks","are","misleading"],"value":"The ARC-AGI-3 scorecard is extremely misleading given that it clearly states itself that "with [the responses API] harness, we estimate Sol would score in the ballpark of ~30%." but it shows a score of 7.8% for GPT-5.6 Sol presumably since if they updated the percentage for GPT-5.6 Sol to the score it would receive with the responses API harness they used for GPT-6 Astra they'd have to do the same for the percentage they show for Opus 5 which would similarly be much higher.
Regardless, the result is still valid as the original benchmark harness is definitely unreasonably handicapped, and if a harness alone can help the LLM saturate the benchmark with a near perfect score then the combination of the two must still be effectively AGI in the sense of passing the most famous benchmark designed specifically to measure AGI progress, after multiple iterations of progressively making it harder.
I think it is fair to say that this is probably effectively AGI if the benchmarks are remotely accurate - even with Fable, I've been at the point personally where I am reasonably confident that there's essentially nothing that I am better than Fable at despite generally being substantively above average on human benchmarks. If Astra's this much better than Fable, I'm ready to call AGI here.
For the many people who resist the AGI label possibly ever being achieved, I'd be curious to hear takes on what would make you think Astra is yet to be AGI, and what would still need to be achieved for this to effectively be AGI from this point forward."},"story_title":{"matchLevel":"none","matchedWords":[],"value":"GPT-6 Astra"},"story_url":{"matchLevel":"none","matchedWords":[],"value":"https://openai.com/index/gpt-6-astra/"}},"_tags":["comment","author_intenex","story_49554643"],"author":"intenex","children":[49556479,49556488,49556523,49556559,49556656,49556692,49556693,49556793,49556885,49556887,49557040,49557066,49557085,49557142,49557156,49557201,49557202,49557388,49557389,49557473,49557485,49557500,49557511,49557547,49557590,49557628,49557672,49557754,49557921,49557980,49558032,49558276,49558336,49558640,49558709,49558795,49558835,49558927,49558971,49559220,49559309,49559360,49559712,49559863,49559881,49560092,49560218,49560404,49560406,49560602,49560621,49560635,49560746,49560790,49560997,49561098,49561132,49561343,49561439,49561809,49562045,49562051],"comment_text":"The ARC-AGI-3 scorecard is extremely misleading given that it clearly states itself that "with [the responses API] harness, we estimate Sol would score in the ballpark of ~30%." but it shows a score of 7.8% for GPT-5.6 Sol presumably since if they updated the percentage for GPT-5.6 Sol to the score it would receive with the responses API harness they used for GPT-6 Astra they'd have to do the same for the percentage they show for Opus 5 which would similarly be much higher.
Regardless, the result is still valid as the original benchmark harness is definitely unreasonably handicapped, and if a harness alone can help the LLM saturate the benchmark with a near perfect score then the combination of the two must still be effectively AGI in the sense of passing the most famous benchmark designed specifically to measure AGI progress, after multiple iterations of progressively making it harder.
I think it is fair to say that this is probably effectively AGI if the benchmarks are remotely accurate - even with Fable, I've been at the point personally where I am reasonably confident that there's essentially nothing that I am better than Fable at despite generally being substantively above average on human benchmarks. If Astra's this much better than Fable, I'm ready to call AGI here.
For the many people who resist the AGI label possibly ever being achieved, I'd be curious to hear takes on what would make you think Astra is yet to be AGI, and what would still need to be achieved for this to effectively be AGI from this point forward.","created_at":"2026-09-03T20:31:08Z","created_at_i":1788467468,"objectID":"49556467","parent_id":49554643,"story_id":49554643,"story_title":"GPT-6 Astra","story_url":"https://openai.com/index/gpt-6-astra/","updated_at":"2026-09-15T03:03:54Z"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"nabakin"},"comment_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["llm","benchmarks","are","misleading"],"value":"Exactly. Falcon-180b had a lot of hype at first but the community soon realized it was nearly worthless. Easily outperformed by smaller LLMs in the general case.
Now they are back and claiming their falcon-11b LLM outperforms Llama 3 8b. I already see a number of issues with this:
- falcon-11b is like 40% larger than Llama 3 8b so how can you compare them when they aren't in the same size class
- their claim seems to be based on automated benchmarks when it has long been clear that automated benchmarks are not enough to make that claim
- some of their automated benchmarks are wildly lower than Llama 3 8b's scores. It only beats Llama 3 8b on one benchmark and just barely. I can make an LLM does the best anyone has ever seen on one benchmark, but that doesn't mean my LLM is good. Far from it
- clickbait headline with knowingly premature claims because there has been zero human evaluation testing
- they claim their LLM is better than Llama 3 but completely ignore Llama 3 70b
Honestly, it annoys me how much attention tiiuae get when they haven't produced anything useful and continue this misleading clickbait."},"story_title":{"matchLevel":"none","matchedWords":[],"value":"Falcon 2"},"story_url":{"matchLevel":"none","matchedWords":[],"value":"https://www.tii.ae/news/falcon-2-uaes-technology-innovation-institute-releases-new-ai-model-series-outperforming-metas"}},"_tags":["comment","author_nabakin","story_40344302"],"author":"nabakin","comment_text":"Exactly. Falcon-180b had a lot of hype at first but the community soon realized it was nearly worthless. Easily outperformed by smaller LLMs in the general case.
Now they are back and claiming their falcon-11b LLM outperforms Llama 3 8b. I already see a number of issues with this:
- falcon-11b is like 40% larger than Llama 3 8b so how can you compare them when they aren't in the same size class
- their claim seems to be based on automated benchmarks when it has long been clear that automated benchmarks are not enough to make that claim
- some of their automated benchmarks are wildly lower than Llama 3 8b's scores. It only beats Llama 3 8b on one benchmark and just barely. I can make an LLM does the best anyone has ever seen on one benchmark, but that doesn't mean my LLM is good. Far from it
- clickbait headline with knowingly premature claims because there has been zero human evaluation testing
- they claim their LLM is better than Llama 3 but completely ignore Llama 3 70b
Honestly, it annoys me how much attention tiiuae get when they haven't produced anything useful and continue this misleading clickbait.","created_at":"2024-05-13T18:13:15Z","created_at_i":1715623995,"objectID":"40346407","parent_id":40344589,"story_id":40344302,"story_title":"Falcon 2","story_url":"https://www.tii.ae/news/falcon-2-uaes-technology-innovation-institute-releases-new-ai-model-series-outperforming-metas","updated_at":"2024-09-20T17:03:07Z"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"PheonixPharts"},"comment_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["llm","benchmarks","are","misleading"],"value":"The benchmarks are so easy to game it's not meaningful unless you see a major difference and even then it can be questionable. For example, did anyone read the Gemini paper[0]?
The team details the slight modification they made to the metric which is basically a tweak on maj@32 that worked in their favor. They choose to do the majority vote of 32 samples unless the probability of the highest sample was too low and then they chose greedy, they also had to learn what "too low" was. Sound convoluted and a bit contrived? They even show plots that clearly demonstrate they settled on the metric because it's the one where GPT-4 results remain they same, but Gemini wins.
Don't get me wrong, LLM evaluation is essential. But in order to do this you need to be very consistent between comparisons and even then the results can be misleading.
0. https://arxiv.org/pdf/2312.11805.pdf"},"story_title":{"fullyHighlighted":false,"matchLevel":"partial","matchedWords":["are"],"value":"Four new models that are benchmarking near or above GPT-4"},"story_url":{"matchLevel":"none","matchedWords":[],"value":"https://simonwillison.net/2024/Mar/8/gpt-4-barrier/"}},"_tags":["comment","author_PheonixPharts","story_39643894"],"author":"PheonixPharts","children":[39645984],"comment_text":"The benchmarks are so easy to game it's not meaningful unless you see a major difference and even then it can be questionable. For example, did anyone read the Gemini paper[0]?
The team details the slight modification they made to the metric which is basically a tweak on maj@32 that worked in their favor. They choose to do the majority vote of 32 samples unless the probability of the highest sample was too low and then they chose greedy, they also had to learn what "too low" was. Sound convoluted and a bit contrived? They even show plots that clearly demonstrate they settled on the metric because it's the one where GPT-4 results remain they same, but Gemini wins.
Don't get me wrong, LLM evaluation is essential. But in order to do this you need to be very consistent between comparisons and even then the results can be misleading.
0. https://arxiv.org/pdf/2312.11805.pdf","created_at":"2024-03-08T18:37:25Z","created_at_i":1709923045,"objectID":"39644363","parent_id":39643894,"story_id":39643894,"story_title":"Four new models that are benchmarking near or above GPT-4","story_url":"https://simonwillison.net/2024/Mar/8/gpt-4-barrier/","updated_at":"2024-09-20T16:37:09Z"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"SirMaster"},"comment_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["llm","benchmarks","are","misleading"],"value":"This is my biggest peeve when people say that LLMs are as capable as humans or that we have achieved AGI or are close or things like that.
But then when I get a subpar result, they always tell me I'm "prompting wrong". LLMs may be very capable of great human level output, but in my experience leave a LOT to be desired in terms of human level understanding of the question or prompt.
I think rating an LLM vs a human or AGI should include it's ability to understand a prompt like a human or like an averagely generally intelligent system should be able to.
Are there any benchmarks on that? Like how well LLMs do with misleading prompts or sparsely quantified prompts compared to one another?
Because if a good prompt is as important as people say, then the model's ability to understand a prompt or perhaps poor prompt could have a massive impact on its output."},"story_title":{"matchLevel":"none","matchedWords":[],"value":"I want to wash my car. The car wash is 50 meters away. Should I walk or drive?"},"story_url":{"matchLevel":"none","matchedWords":[],"value":"https://mastodon.world/@knowmadd/116072773118828295"}},"_tags":["comment","author_SirMaster","story_47031580"],"author":"SirMaster","children":[47041646,47042401],"comment_text":"This is my biggest peeve when people say that LLMs are as capable as humans or that we have achieved AGI or are close or things like that.
But then when I get a subpar result, they always tell me I'm "prompting wrong". LLMs may be very capable of great human level output, but in my experience leave a LOT to be desired in terms of human level understanding of the question or prompt.
I think rating an LLM vs a human or AGI should include it's ability to understand a prompt like a human or like an averagely generally intelligent system should be able to.
Are there any benchmarks on that? Like how well LLMs do with misleading prompts or sparsely quantified prompts compared to one another?
Because if a good prompt is as important as people say, then the model's ability to understand a prompt or perhaps poor prompt could have a massive impact on its output.","created_at":"2026-02-16T17:59:46Z","created_at_i":1771264786,"objectID":"47038052","parent_id":47031882,"story_id":47031580,"story_title":"I want to wash my car. The car wash is 50 meters away. Should I walk or drive?","story_url":"https://mastodon.world/@knowmadd/116072773118828295","updated_at":"2026-03-05T23:33:54Z"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"imiric"},"comment_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["llm","benchmarks","are","misleading"],"value":"Personal experience?
I don't care about benchmarks, nor LLMs' capability to solve puzzles. This is the first thing AI companies optimize their training for, which is misleading and practically false advertising.
I care about how good LLMs are for helping me with specific tasks. Do they generate code that on the surface appears correct, yet on closer inspection has security and performance issues, is unnecessarily complex, often doesn't even compile, which takes me more time to troubleshoot and fix than if I were to write it myself? Do they explain concepts confidently while being wrong, which I have no way of knowing unless I'm a domain expert? Do they repeat all these issues even after careful re-prompting and with all the contextual information they would need? Does all this waste my time more than it helps?
The answer is yes to all of the above.
So while we can argue whether LLMs can think and reason, my personal experience tells me that they absolutely cannot, and that any similarity to what humans can do is nothing but an illusion."},"story_title":{"matchLevel":"none","matchedWords":[],"value":"\"The Illusion of Thinking\" \u2013 Thoughts on This Important Paper"},"story_url":{"matchLevel":"none","matchedWords":[],"value":"https://hardcoresoftware.learningbyshipping.com/p/233-the-illusion-of-thinking-thoughts"}},"_tags":["comment","author_imiric","story_44234626"],"author":"imiric","comment_text":"Personal experience?
I don't care about benchmarks, nor LLMs' capability to solve puzzles. This is the first thing AI companies optimize their training for, which is misleading and practically false advertising.
I care about how good LLMs are for helping me with specific tasks. Do they generate code that on the surface appears correct, yet on closer inspection has security and performance issues, is unnecessarily complex, often doesn't even compile, which takes me more time to troubleshoot and fix than if I were to write it myself? Do they explain concepts confidently while being wrong, which I have no way of knowing unless I'm a domain expert? Do they repeat all these issues even after careful re-prompting and with all the contextual information they would need? Does all this waste my time more than it helps?
The answer is yes to all of the above.
So while we can argue whether LLMs can think and reason, my personal experience tells me that they absolutely cannot, and that any similarity to what humans can do is nothing but an illusion.","created_at":"2025-06-14T10:42:41Z","created_at_i":1749897761,"objectID":"44275535","parent_id":44275119,"story_id":44234626,"story_title":"\"The Illusion of Thinking\" \u2013 Thoughts on This Important Paper","story_url":"https://hardcoresoftware.learningbyshipping.com/p/233-the-illusion-of-thinking-thoughts","updated_at":"2025-06-14T14:26:56Z"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"Swenrekcah"},"comment_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["llm","benchmarks","are","misleading"],"value":"That is not true. Even ChatGPT understands how they are different, I won\u2019t paste the whole response but here are the differences it highlights:
Key differences:
1. Intent and harm:\n \u2022 VW\u2019s actions directly violated laws and had environmental and health consequences. Optimizing LLMs for chess benchmarks, while arguably misleading, doesn\u2019t have immediate real-world harms.\n 2. Scope: Chess-specific optimization is generally a transparent choice within AI research. It\u2019s not a hidden \u201cdefeat device\u201d but rather an explicit design goal.\n 3. Broader impact: LLMs fine-tuned for benchmarks often still retain general-purpose capabilities. They aren\u2019t necessarily \u201cbroken\u201d outside chess, whereas VW cars fundamentally failed to meet emissions standards."},"story_title":{"fullyHighlighted":false,"matchLevel":"partial","matchedWords":["llm"],"value":"Something weird is happening with LLMs and chess"},"story_url":{"matchLevel":"none","matchedWords":[],"value":"https://dynomight.substack.com/p/chess"}},"_tags":["comment","author_Swenrekcah","story_42138289"],"author":"Swenrekcah","comment_text":"That is not true. Even ChatGPT understands how they are different, I won\u2019t paste the whole response but here are the differences it highlights:
Key differences:
1. Intent and harm:\n \u2022 VW\u2019s actions directly violated laws and had environmental and health consequences. Optimizing LLMs for chess benchmarks, while arguably misleading, doesn\u2019t have immediate real-world harms.\n 2. Scope: Chess-specific optimization is generally a transparent choice within AI research. It\u2019s not a hidden \u201cdefeat device\u201d but rather an explicit design goal.\n 3. Broader impact: LLMs fine-tuned for benchmarks often still retain general-purpose capabilities. They aren\u2019t necessarily \u201cbroken\u201d outside chess, whereas VW cars fundamentally failed to meet emissions standards.","created_at":"2024-11-15T14:10:12Z","created_at_i":1731679812,"objectID":"42147078","parent_id":42146749,"story_id":42138289,"story_title":"Something weird is happening with LLMs and chess","story_url":"https://dynomight.substack.com/p/chess","updated_at":"2024-11-15T20:11:08Z"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"cyrilou242"},"comment_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["llm","benchmarks","are","misleading"],"value":"Sadly the paper uses for benchmarks datasets that:\n- are known to be pretty useless\n- contain mistakes \n- can be misleading with a naive F1-score measure. (to be fair they write "we looked at the F1-Score, under which both partial and full anomaly detection are considered correct identification" so this may be mitigated, but it's not clear)
See https://kdd-milets.github.io/milets2021/slides/Irrational%20...\nSo it's hard to take any benchmark from the paper seriously.\nThe paper is also ignoring any recent work (like > 2018) on univariate timeseries anomaly detection in the matrix profile space (eg MADRID).
The "practicality of usage" and conclusion sections are pretty correct though: it's expensive, slow, and no-shot is worthless if some other methods can train and infer in orders of magnitude less time.
It would have been interesting to see how the DETECTOR method performs when the LLM forecasting is replaced with some standard forecasting. (eg some auto ETS, if possible robust to anomalies in the training data). It looks like the natural follow up of this article is to remove the LLM altogether."},"story_title":{"matchLevel":"none","matchedWords":[],"value":"MIT researchers use large language models to flag problems in complex systems"},"story_url":{"matchLevel":"none","matchedWords":[],"value":"https://news.mit.edu/2024/researchers-use-large-language-models-to-flag-problems-0814"}},"_tags":["comment","author_cyrilou242","story_41253544"],"author":"cyrilou242","children":[41254686,41260565],"comment_text":"Sadly the paper uses for benchmarks datasets that:\n- are known to be pretty useless\n- contain mistakes \n- can be misleading with a naive F1-score measure. (to be fair they write "we looked at the F1-Score, under which both partial and full anomaly detection are considered correct identification" so this may be mitigated, but it's not clear)
See https://kdd-milets.github.io/milets2021/slides/Irrational%20...\nSo it's hard to take any benchmark from the paper seriously.\nThe paper is also ignoring any recent work (like > 2018) on univariate timeseries anomaly detection in the matrix profile space (eg MADRID).
The "practicality of usage" and conclusion sections are pretty correct though: it's expensive, slow, and no-shot is worthless if some other methods can train and infer in orders of magnitude less time.
It would have been interesting to see how the DETECTOR method performs when the LLM forecasting is replaced with some standard forecasting. (eg some auto ETS, if possible robust to anomalies in the training data). It looks like the natural follow up of this article is to remove the LLM altogether.","created_at":"2024-08-15T10:02:44Z","created_at_i":1723716164,"objectID":"41254432","parent_id":41253544,"story_id":41253544,"story_title":"MIT researchers use large language models to flag problems in complex systems","story_url":"https://news.mit.edu/2024/researchers-use-large-language-models-to-flag-problems-0814","updated_at":"2024-09-20T17:39:00Z"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"strofcon"},"comment_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["llm","benchmarks","are","misleading"],"value":"I think you make a good point, benchmarks and metrics are indeed a better proxy for performance. Seems worth pointing out that, while "nowhere near half in [your] experience" are completely wrong, I don't take your word for it either. :-)
The trouble in my view is that the only way to know that the answers you're getting are accurate and not misleading is to study up on the answers elsewhere - which is a great habit to nurture, but is also precisely why these tools tend toward uselessness in their "general AI" bids. If I can't know how the answer was built, or how good that answer is, there's no point asking it - I'll just do my own reading and apply appropriate discernment as I go.
To be fair, hardly anyone does this today, nor did they before LLM-based chat bots... So it's a moot point, because society is largely doomed anyway. But a moot point can still be a valid one.
I also think the author makes a good point that we frequently confuse performance for competence. "It does a really good job at <X>!... or at least does a damn fine job of mimicking someone who acts like they do a really good job at <X>!"
By way of analogy, consider Elon Musk - by all appearances, he's a genius and is saving humanity - but by dint of his narcissism and largely smooth-brained approach to... well... everything... he's running all of us into an earlier planet-size grave than is necessary. His performance is fantastic, his competence is nonexistent."},"story_title":{"matchLevel":"none","matchedWords":[],"value":"Rodney Brooks on GPT-4"},"story_url":{"matchLevel":"none","matchedWords":[],"value":"https://spectrum.ieee.org/gpt-4-calm-down"}},"_tags":["comment","author_strofcon","story_36017309"],"author":"strofcon","children":[36017756],"comment_text":"I think you make a good point, benchmarks and metrics are indeed a better proxy for performance. Seems worth pointing out that, while "nowhere near half in [your] experience" are completely wrong, I don't take your word for it either. :-)
The trouble in my view is that the only way to know that the answers you're getting are accurate and not misleading is to study up on the answers elsewhere - which is a great habit to nurture, but is also precisely why these tools tend toward uselessness in their "general AI" bids. If I can't know how the answer was built, or how good that answer is, there's no point asking it - I'll just do my own reading and apply appropriate discernment as I go.
To be fair, hardly anyone does this today, nor did they before LLM-based chat bots... So it's a moot point, because society is largely doomed anyway. But a moot point can still be a valid one.
I also think the author makes a good point that we frequently confuse performance for competence. "It does a really good job at <X>!... or at least does a damn fine job of mimicking someone who acts like they do a really good job at <X>!"
By way of analogy, consider Elon Musk - by all appearances, he's a genius and is saving humanity - but by dint of his narcissism and largely smooth-brained approach to... well... everything... he's running all of us into an earlier planet-size grave than is necessary. His performance is fantastic, his competence is nonexistent.","created_at":"2023-05-21T02:52:56Z","created_at_i":1684637576,"objectID":"36017609","parent_id":36017443,"story_id":36017309,"story_title":"Rodney Brooks on GPT-4","story_url":"https://spectrum.ieee.org/gpt-4-calm-down","updated_at":"2024-09-20T14:08:26Z"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"eis"},"comment_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["llm","benchmarks","are","misleading"],"value":"> Increases in LLM scores are now mainly driven by post training (evidence in next section) and are probably a function of amount of synthetic data. They are learning to solve ARC tasks, not learn general abstract reasoning
Agreed and that's for any benchmark. Private tests are better but you still have to trust the provider to not log and use them for training.
That's why I like when a new set of tests like a new ARC-AGI version is published, that's where you can see which of the models abstracted to more general capabilities instead of being focused on the previous tasks. Most models completely fail new ARC-AGI tests.
The "67 cents" part though is misleading imho. You can't extrapolate from there and think that investing say $100 will get you a lot better results. You hit a ceiling very fast and investing into more compute will give you diminishing results. So yes, you can train a custom model to do somewhat decently on a specific set of tasks but then what?"},"story_title":{"matchLevel":"none","matchedWords":[],"value":"I trained a small transformer in 1.5hrs and it beats many LLMs"},"story_url":{"matchLevel":"none","matchedWords":[],"value":"https://mvakde.github.io/blog/44-on-arc-1/"}},"_tags":["comment","author_eis","story_49519939"],"author":"eis","children":[49520241,49520741,49525435],"comment_text":"> Increases in LLM scores are now mainly driven by post training (evidence in next section) and are probably a function of amount of synthetic data. They are learning to solve ARC tasks, not learn general abstract reasoning
Agreed and that's for any benchmark. Private tests are better but you still have to trust the provider to not log and use them for training.
That's why I like when a new set of tests like a new ARC-AGI version is published, that's where you can see which of the models abstracted to more general capabilities instead of being focused on the previous tasks. Most models completely fail new ARC-AGI tests.
The "67 cents" part though is misleading imho. You can't extrapolate from there and think that investing say $100 will get you a lot better results. You hit a ceiling very fast and investing into more compute will give you diminishing results. So yes, you can train a custom model to do somewhat decently on a specific set of tasks but then what?","created_at":"2026-09-01T10:40:25Z","created_at_i":1788259225,"objectID":"49520167","parent_id":49519939,"story_id":49519939,"story_title":"I trained a small transformer in 1.5hrs and it beats many LLMs","story_url":"https://mvakde.github.io/blog/44-on-arc-1/","updated_at":"2026-09-03T20:36:18Z"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"ftxbro"},"comment_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["llm","benchmarks","are","misleading"],"value":"the most misleading things i saw in llm ai were the reports of performances of local models. several independent reports of various models were saying better than gpt-3.5 or even gpt-4 and it was never true. maybe they are doing their benchmarks in a silly way. it made me not trust these analyses of local models.
EDIT: everyone in that reddit calling them out for this. i didn't realize it was such a meme lol"},"story_title":{"matchLevel":"none","matchedWords":[],"value":"WizardLM-30B v1.0: beats Guanaco-65B"},"story_url":{"matchLevel":"none","matchedWords":[],"value":"https://old.reddit.com/r/LocalLLaMA/comments/142iw20/official_wizardlm30b_v10_released_can_beat/"}},"_tags":["comment","author_ftxbro","story_36216650"],"author":"ftxbro","comment_text":"the most misleading things i saw in llm ai were the reports of performances of local models. several independent reports of various models were saying better than gpt-3.5 or even gpt-4 and it was never true. maybe they are doing their benchmarks in a silly way. it made me not trust these analyses of local models.
EDIT: everyone in that reddit calling them out for this. i didn't realize it was such a meme lol","created_at":"2023-06-06T18:10:01Z","created_at_i":1686075001,"objectID":"36216830","parent_id":36216650,"story_id":36216650,"story_title":"WizardLM-30B v1.0: beats Guanaco-65B","story_url":"https://old.reddit.com/r/LocalLLaMA/comments/142iw20/official_wizardlm30b_v10_released_can_beat/","updated_at":"2024-09-20T14:11:05Z"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"famouswaffles"},"comment_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["llm","benchmarks","are","misleading"],"value":">If the benchmark is not representative of normal usage* then the benchmark and the plot being shown are not useful at all from a user/business perspective and the focus on the breakthrough scores of o3-low and o3-high in ARC-AGI would be highly misleading.
ARC is a very hyped benchmark in the industry so letting us know the results is something any company would do whether it had a direct representation on normal usage or not.
>Even if it is pure marketing they expect people to draw conclusions using the perf/cost plot from Arc.
Again, people care about ARC, they don't care doing the things ARC questions ask. That it is un-economical to pay the price to use o3 for ARC does not mean it would be un-economical to do so for the tasks people actually want to use LLMs for. What does 3x the performance in say coding mean? You really think companies/users wouldn't put up with the increased price for that? You think they have Mturkers to turn to like they do with ARC?
ARC is literally the quintessential 'easy for humans, hard for ai' benchmark. Even if you discard the 'difficulty to price won't scale the same' argument, it makes no sense to use it for an economics comparison."},"story_title":{"matchLevel":"none","matchedWords":[],"value":"GPT-5 is behind schedule"},"story_url":{"matchLevel":"none","matchedWords":[],"value":"https://www.wsj.com/tech/ai/openai-gpt5-orion-delays-639e7693"}},"_tags":["comment","author_famouswaffles","story_42485938"],"author":"famouswaffles","children":[42492129],"comment_text":">If the benchmark is not representative of normal usage* then the benchmark and the plot being shown are not useful at all from a user/business perspective and the focus on the breakthrough scores of o3-low and o3-high in ARC-AGI would be highly misleading.
ARC is a very hyped benchmark in the industry so letting us know the results is something any company would do whether it had a direct representation on normal usage or not.
>Even if it is pure marketing they expect people to draw conclusions using the perf/cost plot from Arc.
Again, people care about ARC, they don't care doing the things ARC questions ask. That it is un-economical to pay the price to use o3 for ARC does not mean it would be un-economical to do so for the tasks people actually want to use LLMs for. What does 3x the performance in say coding mean? You really think companies/users wouldn't put up with the increased price for that? You think they have Mturkers to turn to like they do with ARC?
ARC is literally the quintessential 'easy for humans, hard for ai' benchmark. Even if you discard the 'difficulty to price won't scale the same' argument, it makes no sense to use it for an economics comparison.","created_at":"2024-12-23T03:42:33Z","created_at_i":1734925353,"objectID":"42491537","parent_id":42491089,"story_id":42485938,"story_title":"GPT-5 is behind schedule","story_url":"https://www.wsj.com/tech/ai/openai-gpt5-orion-delays-639e7693","updated_at":"2025-11-10T20:40:58Z"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"Kim_Bruning"},"comment_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["llm","benchmarks","are","misleading"],"value":"Wait a second, none of it? How about formal reasoning? Regular IF-THEN-ELSE can do simple logic, and prolog can do inference already. So are you saying LLMs can't do stuff that computers have been doing for ages?
To test this for some of my own uses, I've had this quick benchmark with progressively harder reasoning needed to understand novel prose. Each generation of models I've tested can unravel more layers of deliberately misleading writing; while meanwhile I've seen humans give up on the first question.
So either the models are applying reasoning, or some form of magic is happening."},"story_title":{"matchLevel":"none","matchedWords":[],"value":"When did Google get so weird?"},"story_url":{"matchLevel":"none","matchedWords":[],"value":"https://sancho.bearblog.dev/google-weird/"}},"_tags":["comment","author_Kim_Bruning","story_49870367"],"author":"Kim_Bruning","comment_text":"Wait a second, none of it? How about formal reasoning? Regular IF-THEN-ELSE can do simple logic, and prolog can do inference already. So are you saying LLMs can't do stuff that computers have been doing for ages?
To test this for some of my own uses, I've had this quick benchmark with progressively harder reasoning needed to understand novel prose. Each generation of models I've tested can unravel more layers of deliberately misleading writing; while meanwhile I've seen humans give up on the first question.
So either the models are applying reasoning, or some form of magic is happening.","created_at":"2026-09-28T07:04:11Z","created_at_i":1790579051,"objectID":"49874541","parent_id":49873492,"story_id":49870367,"story_title":"When did Google get so weird?","story_url":"https://sancho.bearblog.dev/google-weird/","updated_at":"2026-09-28T17:47:05Z"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"Nezk"},"comment_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["llm","benchmarks","are","misleading"],"value":"And benchmarking this language against Isabelle/Agda/Lean/Rocq is strange. The time taken for those systems to perform their checks is mostly spent on elaboration, which includes unification against metavariables, typeclass resolution and tactics. Bend has none of that (there are no type classes or traits, and according to the README, everything must be fully annotated and nothing inferred). This means that the benchmark is comparing Bend's checker to the other systems' elaborators + kernels rather than their kernels (Agda doesn't have this separation though). The latter would be a fairer comparison, and in this area the other systems are already fast. Framing it as "outperforming every proof assistant" without that caveat is misleading.
There is also a problem with LAWS.bend. The typechecker only guarantees that your code satisfies what's written in LAWS.bend \u2014 not that LAWS.bend says what you actually meant. There is nothing to stop an LLM from "satisfying" a law with a vacuous or narrower-than-intended formalisation \u2014 the trust problem simply shifts from the code to the specification (which could be also generated by LLM, and therefore incorrect). The repository even admits that the compiler itself is 99% LLM generated and not yet fully audited, which seems a questionable basis on which to build a "mathematical guarantees" marketing."},"story_title":{"matchLevel":"none","matchedWords":[],"value":"Bend \u2013 A language that blocks AI mistakes via proof, on CPU and GPU"},"story_url":{"matchLevel":"none","matchedWords":[],"value":"https://bend-lang.com/"}},"_tags":["comment","author_Nezk","story_49746163"],"author":"Nezk","comment_text":"And benchmarking this language against Isabelle/Agda/Lean/Rocq is strange. The time taken for those systems to perform their checks is mostly spent on elaboration, which includes unification against metavariables, typeclass resolution and tactics. Bend has none of that (there are no type classes or traits, and according to the README, everything must be fully annotated and nothing inferred). This means that the benchmark is comparing Bend's checker to the other systems' elaborators + kernels rather than their kernels (Agda doesn't have this separation though). The latter would be a fairer comparison, and in this area the other systems are already fast. Framing it as "outperforming every proof assistant" without that caveat is misleading.
There is also a problem with LAWS.bend. The typechecker only guarantees that your code satisfies what's written in LAWS.bend \u2014 not that LAWS.bend says what you actually meant. There is nothing to stop an LLM from "satisfying" a law with a vacuous or narrower-than-intended formalisation \u2014 the trust problem simply shifts from the code to the specification (which could be also generated by LLM, and therefore incorrect). The repository even admits that the compiler itself is 99% LLM generated and not yet fully audited, which seems a questionable basis on which to build a "mathematical guarantees" marketing.","created_at":"2026-09-18T08:50:43Z","created_at_i":1789721443,"objectID":"49751733","parent_id":49750750,"story_id":49746163,"story_title":"Bend \u2013 A language that blocks AI mistakes via proof, on CPU and GPU","story_url":"https://bend-lang.com/","updated_at":"2026-09-18T14:19:49Z"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"LoadingALIAS"},"comment_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["llm","benchmarks","are","misleading"],"value":"Karolin, I\u2019ll answer this directly, and try to stay on track, because several claims here are materially wrong.
First things first, I am not claiming I've had any third-party audit; I'm not claiming I've been FIPS validated, or telling anyone \u201cuse this blindly in high-assurance production.\u201d If that is your bar, then yes: do not use rscrypto yet. That is a fair warning... and it's plastered everywhere. I almost feel like you didn't read the codebase, the thread here, etc.
Now, to your constant-time criticism - it's just not correct. It's wrong. It's misleading. It shouldn't come from someone leading Rosenpass.
\u201cNot a global constant-time claim\u201d does not mean \u201csecret-bearing cryptographic ops branch on secrets.\u201d I've said this twice now - verbatim. It means I am refusing to make a fake blanket claim over APIs where constant-time behavior is either irrelevant, impossible, or the wrong security property.
DER parsing is not supposed to be constant-time over DER structure. PHC string decoding is not supposed to be constant-time over ASCII syntax. OS randomness is not a constant-time primitive. RSA public verification/encryption operates on public inputs. Rejection of public length/shape before entering a primitive boundary is NOT a secret-dependent timing leak. Argon2d intentionally has data-dependent memory access; treating it as a blanket CT primitive would be wrong - it's a lie at that point.
That list was a boundary, Karolin, not an admission of insecure primitives.
The MAC/hash point is also mixing categories. A MAC verification claim is about the keyed construction and verification surface: no secret-dependent behavior in the keyed path that matters for the construction, opaque failure, and constant-time tag comparison. An unkeyed hash API as a whole does not get the same global claim because message length, streaming shape, finalization shape, feature-dispatched kernels, and non-crypto hash/checksum APIs are public-input machinery. \u201cHMAC uses a hash\u201d does not imply \u201cthe entire hash API must be globally constant-time under every possible use.\u201d
On \u201calgo/profile negotiation\u201d: this is not TLS-style negotiation. It is closed protocol identifier mapping for RSA profiles: JWT/COSE/TLS/X.509 identifiers are parsed into EXPLICIT supported profile enums, and unsupported/confused algos are rejected. If the word \u201cnegotiation\u201d suggested downgrade-prone protocol behavior, I should fix the language, sure, but the feature is not an open negotiation mechanism.
With respect to the perf, if you believe a benchmark win comes from a specific missing constant-time property, name the primitive, input class, backend, and compared impl. \u201cIt might be faster because it is insecure\u201d is not evidence... it's negativity and hyperbole.
Also, every single line of this codebase is my responsibility. My usage of an LLM is not a substitute for review, tests, vectors, fuzzing, Miri, and/or a third-party audit. If you're insinuating that using an LLM today is a vulnerability report - I'm afraid you're falling behind the best engineers in the world and are sorely mistaken.
Finally, and this is the one that really gets under my skin, which is likely your goal given your very public history of such... my putting code in public doesn't force cryptographers to review it. It invites scrutiny. If there is duplication that materially increases audit cost, point to it and I will remove or justify it. If there is a side channel, GIVE ME THE PATH and I will treat it as a security issue. But \u201cdo not publish until the community has reviewed it\u201d is backwards: review cannot happen against code that is not visible. You're plastered all across the Internet championing this exact maxim, are you not? Did I not read an interview this morning where you went on about the importance of OSS crypto?
Please, take some time to review the codebase or don't; that's up to you... but don't come in here and trash my work improving and unifying well known primitives. Also, this isn't a whimsical codebase vibe coded by Claude overnight. This is a year of my life reading, understanding, and improving the inefficiencies in primitives I've used for years.
The entire point of a pre-v1 review by the community is the same exact thing OSS engineers/contributors have done for decades. I'm looking for the community to point out glaring issues, errors, API shapes, code smells, etc. You came in here with a bunch of incorrect claims and emotion. This codebase isn't perfect, but I assure you - it will be, and it will lead Rust by default becasue there simple will not be a better option.
Happy to talk one on one, or we could put something together where others have a chance to chime in. Either way, have a good one."},"story_title":{"matchLevel":"none","matchedWords":[],"value":"Show HN: Rscrypto, pure-Rust crypto with industry leading public benches"},"story_url":{"matchLevel":"none","matchedWords":[],"value":"https://github.com/loadingalias/rscrypto"}},"_tags":["comment","author_LoadingALIAS","story_48386343"],"author":"LoadingALIAS","comment_text":"Karolin, I\u2019ll answer this directly, and try to stay on track, because several claims here are materially wrong.
First things first, I am not claiming I've had any third-party audit; I'm not claiming I've been FIPS validated, or telling anyone \u201cuse this blindly in high-assurance production.\u201d If that is your bar, then yes: do not use rscrypto yet. That is a fair warning... and it's plastered everywhere. I almost feel like you didn't read the codebase, the thread here, etc.
Now, to your constant-time criticism - it's just not correct. It's wrong. It's misleading. It shouldn't come from someone leading Rosenpass.
\u201cNot a global constant-time claim\u201d does not mean \u201csecret-bearing cryptographic ops branch on secrets.\u201d I've said this twice now - verbatim. It means I am refusing to make a fake blanket claim over APIs where constant-time behavior is either irrelevant, impossible, or the wrong security property.
DER parsing is not supposed to be constant-time over DER structure. PHC string decoding is not supposed to be constant-time over ASCII syntax. OS randomness is not a constant-time primitive. RSA public verification/encryption operates on public inputs. Rejection of public length/shape before entering a primitive boundary is NOT a secret-dependent timing leak. Argon2d intentionally has data-dependent memory access; treating it as a blanket CT primitive would be wrong - it's a lie at that point.
That list was a boundary, Karolin, not an admission of insecure primitives.
The MAC/hash point is also mixing categories. A MAC verification claim is about the keyed construction and verification surface: no secret-dependent behavior in the keyed path that matters for the construction, opaque failure, and constant-time tag comparison. An unkeyed hash API as a whole does not get the same global claim because message length, streaming shape, finalization shape, feature-dispatched kernels, and non-crypto hash/checksum APIs are public-input machinery. \u201cHMAC uses a hash\u201d does not imply \u201cthe entire hash API must be globally constant-time under every possible use.\u201d
On \u201calgo/profile negotiation\u201d: this is not TLS-style negotiation. It is closed protocol identifier mapping for RSA profiles: JWT/COSE/TLS/X.509 identifiers are parsed into EXPLICIT supported profile enums, and unsupported/confused algos are rejected. If the word \u201cnegotiation\u201d suggested downgrade-prone protocol behavior, I should fix the language, sure, but the feature is not an open negotiation mechanism.
With respect to the perf, if you believe a benchmark win comes from a specific missing constant-time property, name the primitive, input class, backend, and compared impl. \u201cIt might be faster because it is insecure\u201d is not evidence... it's negativity and hyperbole.
Also, every single line of this codebase is my responsibility. My usage of an LLM is not a substitute for review, tests, vectors, fuzzing, Miri, and/or a third-party audit. If you're insinuating that using an LLM today is a vulnerability report - I'm afraid you're falling behind the best engineers in the world and are sorely mistaken.
Finally, and this is the one that really gets under my skin, which is likely your goal given your very public history of such... my putting code in public doesn't force cryptographers to review it. It invites scrutiny. If there is duplication that materially increases audit cost, point to it and I will remove or justify it. If there is a side channel, GIVE ME THE PATH and I will treat it as a security issue. But \u201cdo not publish until the community has reviewed it\u201d is backwards: review cannot happen against code that is not visible. You're plastered all across the Internet championing this exact maxim, are you not? Did I not read an interview this morning where you went on about the importance of OSS crypto?
Please, take some time to review the codebase or don't; that's up to you... but don't come in here and trash my work improving and unifying well known primitives. Also, this isn't a whimsical codebase vibe coded by Claude overnight. This is a year of my life reading, understanding, and improving the inefficiencies in primitives I've used for years.
The entire point of a pre-v1 review by the community is the same exact thing OSS engineers/contributors have done for decades. I'm looking for the community to point out glaring issues, errors, API shapes, code smells, etc. You came in here with a bunch of incorrect claims and emotion. This codebase isn't perfect, but I assure you - it will be, and it will lead Rust by default becasue there simple will not be a better option.
Happy to talk one on one, or we could put something together where others have a chance to chime in. Either way, have a good one.","created_at":"2026-06-04T14:51:20Z","created_at_i":1780584680,"objectID":"48399569","parent_id":48394475,"story_id":48386343,"story_title":"Show HN: Rscrypto, pure-Rust crypto with industry leading public benches","story_url":"https://github.com/loadingalias/rscrypto","updated_at":"2026-06-05T14:16:00Z"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"famouswaffles"},"comment_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["llm","benchmarks","are","misleading"],"value":">I'm not expecting novel scientific theories today. What I am expecting are signs and hints of such genius.
Like I said, what exactly would you be expecting to see with the capabilities that exist today ? It's not a gotcha, it's a genuine question.
>And, I'm sorry, I don't buy the excuse that these tools are not "yet" as capable as the best and brightest humans.
There's nothing to buy or not buy. They simply aren't. They are unable to do a lot of the things these people do. You can't slot an LLM in place of most knowledge workers and expect everything to be fine and dandy. There's no ambiguity on that.
>They contain the sum of human knowledge, far more than any individual human in history.
It's not really the total sum of human knowledge but let's set that aside. Yeah so ? Einstein, Newton, Von Newman. None of these guys were privy to some super secret knowledge their contemporaries weren't so it's obviously not simply a matter of more knowledge.
>Are they not intelligent, capable of thinking and reasoning?
Yeah they are. And so are humans. So were the peers of all those guys. So why are only a few able to see the next step ? It's not just about knowledge, and intelligence lives in degrees/is a gradient.
>If all this is true, surely we should be seeing incredible results produced by this technology. If not by itself, then surely by "amplifying" the work of the best and brightest humans.
Yeah and that exists. Terence Tao has shared a lot of his (and his peers) experiences on the matter.
https://mathstodon.xyz/@tao/115306424727150237
https://mathstodon.xyz/@tao/115420236285085121
https://mathstodon.xyz/@tao/115416208975810074
>And yet... All we have to show for it are some very good applications of pattern matching and statistics, a bunch of gamed and misleading benchmarks and leaderboards, a whole lot of tech demos, solutions in search of a problem, and the very real problem of flooding us with even more spam, scams, disinformation, and devaluing human work with low-effort garbage.
Well it's a good thing that's not true then"},"story_title":{"matchLevel":"none","matchedWords":[],"value":"A new Google model is nearly perfect on automated handwriting recognition"},"story_url":{"matchLevel":"none","matchedWords":[],"value":"https://generativehistory.substack.com/p/has-google-quietly-solved-two-of"}},"_tags":["comment","author_famouswaffles","story_45887262"],"author":"famouswaffles","children":[45936418],"comment_text":">I'm not expecting novel scientific theories today. What I am expecting are signs and hints of such genius.
Like I said, what exactly would you be expecting to see with the capabilities that exist today ? It's not a gotcha, it's a genuine question.
>And, I'm sorry, I don't buy the excuse that these tools are not "yet" as capable as the best and brightest humans.
There's nothing to buy or not buy. They simply aren't. They are unable to do a lot of the things these people do. You can't slot an LLM in place of most knowledge workers and expect everything to be fine and dandy. There's no ambiguity on that.
>They contain the sum of human knowledge, far more than any individual human in history.
It's not really the total sum of human knowledge but let's set that aside. Yeah so ? Einstein, Newton, Von Newman. None of these guys were privy to some super secret knowledge their contemporaries weren't so it's obviously not simply a matter of more knowledge.
>Are they not intelligent, capable of thinking and reasoning?
Yeah they are. And so are humans. So were the peers of all those guys. So why are only a few able to see the next step ? It's not just about knowledge, and intelligence lives in degrees/is a gradient.
>If all this is true, surely we should be seeing incredible results produced by this technology. If not by itself, then surely by "amplifying" the work of the best and brightest humans.
Yeah and that exists. Terence Tao has shared a lot of his (and his peers) experiences on the matter.
https://mathstodon.xyz/@tao/115306424727150237
https://mathstodon.xyz/@tao/115420236285085121
https://mathstodon.xyz/@tao/115416208975810074
>And yet... All we have to show for it are some very good applications of pattern matching and statistics, a bunch of gamed and misleading benchmarks and leaderboards, a whole lot of tech demos, solutions in search of a problem, and the very real problem of flooding us with even more spam, scams, disinformation, and devaluing human work with low-effort garbage.
Well it's a good thing that's not true then","created_at":"2025-11-15T03:49:28Z","created_at_i":1763178568,"objectID":"45934910","parent_id":45934710,"story_id":45887262,"story_title":"A new Google model is nearly perfect on automated handwriting recognition","story_url":"https://generativehistory.substack.com/p/has-google-quietly-solved-two-of","updated_at":"2026-03-05T22:59:48Z"}],"hitsPerPage":20,"nbHits":27,"nbPages":2,"page":0,"params":"query=LLM+benchmarks+are+misleading&advancedSyntax=true&analyticsTags=backend","processingTimeMS":26,"processingTimingsMS":{"_request":{"roundTrip":22},"afterFetch":{"format":{"highlighting":2,"total":2},"merge":{"mergeLoop":{"prepareNextHit":1,"total":1},"total":1},"total":1},"fetch":{"query":12,"scanning":10,"total":23},"total":26},"query":"LLM benchmarks are misleading","serverTimeMS":29}