Not sure if GLM-4.5 Air is good, but non-Air one is fabulous. I know for free API access there is pollinations ai project. Also llm7. If you just use the web chat's you can use most of the best models for free without API. There are ways to 'emulate' an API automatically.. I was thinking about adding this to my aicodeprep-gui app so it could automatically paste and then cut. Some MCP servers exist that you can use and it will automatically paste or cut from those web chat's and route it to an API interface.
OpenAI offers free tokens for most models, 2.5mil or 250k depending on model. Cerebras has some free limits, Gemini... Meta has plentiful free API for Llama 4 because.. lets face it, it sucks, but it is okay/not bad for stuff like summarizing text.
If you really wanted to code for exactly $0 you could use pollinations ai, in Cline extension (for VS Code) set to use "openai-large" (which is GPT 4.1). If you plan using all the best web chat's like Kimi K2, z.ai's GLM models, Qwen 3 chat, Gemini in AI Studio, OpenAI playground with o3 or o4-mini. You can go forever without being charged money. Pollinations 'openai-large' works fine in Cline as an agent to edit files for you etc."},"story_title":{"fullyHighlighted":false,"matchLevel":"partial","matchedWords":["ai"],"value":"How I code with AI on a budget/free"},"story_url":{"matchLevel":"none","matchedWords":[],"value":"https://wuu73.org/blog/aiguide1.html"}},"_tags":["comment","author_radio879","story_44850913"],"author":"radio879","children":[44852483,44852945],"comment_text":"Well, if you download Qwen Code https://github.com/QwenLM/qwen-code it is free up to 2000 api calls a day.
Not sure if GLM-4.5 Air is good, but non-Air one is fabulous. I know for free API access there is pollinations ai project. Also llm7. If you just use the web chat's you can use most of the best models for free without API. There are ways to 'emulate' an API automatically.. I was thinking about adding this to my aicodeprep-gui app so it could automatically paste and then cut. Some MCP servers exist that you can use and it will automatically paste or cut from those web chat's and route it to an API interface.
OpenAI offers free tokens for most models, 2.5mil or 250k depending on model. Cerebras has some free limits, Gemini... Meta has plentiful free API for Llama 4 because.. lets face it, it sucks, but it is okay/not bad for stuff like summarizing text.
If you really wanted to code for exactly $0 you could use pollinations ai, in Cline extension (for VS Code) set to use "openai-large" (which is GPT 4.1). If you plan using all the best web chat's like Kimi K2, z.ai's GLM models, Qwen 3 chat, Gemini in AI Studio, OpenAI playground with o3 or o4-mini. You can go forever without being charged money. Pollinations 'openai-large' works fine in Cline as an agent to edit files for you etc.","created_at":"2025-08-10T02:39:50Z","created_at_i":1754793590,"objectID":"44852362","parent_id":44851937,"story_id":44850913,"story_title":"How I code with AI on a budget/free","story_url":"https://wuu73.org/blog/aiguide1.html","updated_at":"2026-03-05T22:31:42Z"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"solid_fuel"},"comment_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["which","ai","model","is","best","for","summarizing"],"value":"I keep finding myself contemplating the proscription of AI in the universe of Dune. While the later prequels by Brian Herbert write a fairly typically backstory of an AI rebellion and war, the original novels hint at something - IMO - far more interesting: AI was used by other people to manipulate and control, it didn't take control directly. People rebelled over the sheer amount of power that AI and computers enabled a few people to wield over society and the way it turned human life into one of being a cog in a machine, leading regimented and structured lives.
To wit: while ChatGPT and Gemini and most of these current models are fairly well behaved, when you use a model even for something seemingly innocuous like summarizing an article or an email you are indirectly allowing another person to decide what is important to you. Consider the power that gives other people over you. It pays to put on the devils cap sometimes and imagine a future where the LLMs that power our tools are controlled by people who don't exercise any restraint.
We have already seen shades of this with the amount of influence Facebook and TikTok and Twitter can wield over political discourse, picking which issues are winners and which are losers by choosing (even indirectly, by simply reacting to engagement metrics) what to emphasize and what to suppress. LLMs unlock another level entirely. An LLM can easily summarize an article while conveniently leaving out any negative mentions of certain politicians or parties. They can summarize emails and texts from family and friends while eliding any section asking for help or action.
While most people are somewhat distrustful of obviously biased sources, they don't regard LLMs with the same suspicion. An LLM could easily write a summary of Alan Turing's life while dropping all the bits about his sexuality and the way he was persecuted and castrated by the British government simply for trying to love.
I am not alleging that any of these things have happened, yet. But it is best to think of LLMs and generative AI in general as a tool that works _for someone else_. They can be very useful tools, but they can also be subverted and manipulated in subtle ways, and should not automatically be regarded as unbiased."},"story_title":{"fullyHighlighted":false,"matchLevel":"partial","matchedWords":["ai","is"],"value":"AI is Anti-Human (and assorted qualifications)"},"story_url":{"matchLevel":"none","matchedWords":[],"value":"https://njump.me/naddr1qqxnzde58yerxv3exycrsdpjqgsf03c2gsmx5ef4c9zmxvlew04gdh7u94afnknp33qvv3c94kvwxgsrqsqqqa28nmz2vk"}},"_tags":["comment","author_solid_fuel","story_44428731"],"author":"solid_fuel","children":[44435177],"comment_text":"I keep finding myself contemplating the proscription of AI in the universe of Dune. While the later prequels by Brian Herbert write a fairly typically backstory of an AI rebellion and war, the original novels hint at something - IMO - far more interesting: AI was used by other people to manipulate and control, it didn't take control directly. People rebelled over the sheer amount of power that AI and computers enabled a few people to wield over society and the way it turned human life into one of being a cog in a machine, leading regimented and structured lives.
To wit: while ChatGPT and Gemini and most of these current models are fairly well behaved, when you use a model even for something seemingly innocuous like summarizing an article or an email you are indirectly allowing another person to decide what is important to you. Consider the power that gives other people over you. It pays to put on the devils cap sometimes and imagine a future where the LLMs that power our tools are controlled by people who don't exercise any restraint.
We have already seen shades of this with the amount of influence Facebook and TikTok and Twitter can wield over political discourse, picking which issues are winners and which are losers by choosing (even indirectly, by simply reacting to engagement metrics) what to emphasize and what to suppress. LLMs unlock another level entirely. An LLM can easily summarize an article while conveniently leaving out any negative mentions of certain politicians or parties. They can summarize emails and texts from family and friends while eliding any section asking for help or action.
While most people are somewhat distrustful of obviously biased sources, they don't regard LLMs with the same suspicion. An LLM could easily write a summary of Alan Turing's life while dropping all the bits about his sexuality and the way he was persecuted and castrated by the British government simply for trying to love.
I am not alleging that any of these things have happened, yet. But it is best to think of LLMs and generative AI in general as a tool that works _for someone else_. They can be very useful tools, but they can also be subverted and manipulated in subtle ways, and should not automatically be regarded as unbiased.","created_at":"2025-07-01T00:36:19Z","created_at_i":1751330179,"objectID":"44429429","parent_id":44428731,"story_id":44428731,"story_title":"AI is Anti-Human (and assorted qualifications)","story_url":"https://njump.me/naddr1qqxnzde58yerxv3exycrsdpjqgsf03c2gsmx5ef4c9zmxvlew04gdh7u94afnknp33qvv3c94kvwxgsrqsqqqa28nmz2vk","updated_at":"2025-07-01T15:55:04Z"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"rcody"},"story_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["which","ai","model","is","best","for","summarizing"],"value":"Hi HN, we're Roya and Michael, co-founders of Opusense AI (https://www.opusense.com/), a tool to help engineers and consultants automatically generate construction site reports from typed or voice notes, plus photos.
Here\u2019s a video: https://www.youtube.com/watch?v=u3Pi1iih1_Y.
Before this, I (Roya) worked in human-machine interaction at Huawei, and before that as a construction site inspector for civil engineering firms. I have a PhD in Civil Engineering, and in my experience reporting was by far the most tedious and mind-numbing part of the job.
You\u2019d walk around a site all day taking short notes (maybe, often you'd rely on memory) and snapping photos, then go to three more sites before finally making it back to the office and try to remember everything you wanted to write. Sometimes you\u2019d fill in gaps from memory or you\u2019d keep it purposefully vague. Reports had to be consistent, branded, and checked by senior engineers. It was a huge time sink across the team.
Writing reports was the worst part of the job, so we built Opusense to get rid of it. On-site, users type or dictate short notes (e.g. \u201crebar exposed east end of slab\u201d), and the tool turns them into full sentences, paragraphs, tables, or photo captions in a report template that matches the firm\u2019s format. You can work offline, and it syncs automatically when back online.
Most inspection and reporting tools are built for checklist-style workflows (which is great for home inspections or punch lists), but civil, structural, environmental, or geotechnical engineers usually need freeform notes, not radio buttons.
This is a particularly good fit for LLMs because engineering field reports live in a constrained, conventional domain: similar language, repeated structures, and highly standardized content across firms and projects. There\u2019s a lot of redundancy and grunt work, summarizing the same site conditions, formatting repetitive data, translating field notes into polished paragraphs, all of which LLMs handle well with the right prompting and guardrails. We\u2019re not generating arbitrary prose; we\u2019re transforming structured inputs (notes, images, forms) into structured outputs, with firm-defined templates and required fields that minimize the risk of hallucination. When facts matter (e.g. test results or measurements), we keep them grounded in the user\u2019s input, the model doesn\u2019t invent data because there\u2019s nothing for it to invent. This makes it one of those cases where LLMs aren\u2019t just a novelty, they're genuinely the best tool for the job.
Under the hood, we use a combination of prompt-engineered LLMs and firm-specific formatting rules to get outputs that don\u2019t just sound good, but also look right. We\u2019ve recently added translation features, and we\u2019re iterating quickly based on field feedback. We charge per seat and are deployed at mid size firms, and trialing with some multinational engineering firms who have thousands of reports to file each week. We're also starting to see interest from construction managers and developers who do their own internal QA reporting.
We don't have a self-serve way to try out the product yet, because the way our business works requires templates to be customized by company. But there\u2019s a demo at https://www.youtube.com/watch?v=u3Pi1iih1_Y, and if you want to poke around the UI yourself, here\u2019s a sample account to log in with:
login: hndemo@opusense.com\n password: OpusenseHacker2025\nWe\u2019d love to hear how others are thinking about tools for field work, reporting, or similar workflows (engineering, architectural, etc.). If you\u2019ve built in this space, or have thoughts on how to improve it, we\u2019re all ears!"},"title":{"fullyHighlighted":false,"matchLevel":"partial","matchedWords":["ai","for"],"value":"Launch HN: Opusense (YC X25) \u2013 AI assistant for construction inspectors on site"}},"_tags":["story","author_rcody","story_44042791","launch_hn"],"author":"rcody","children":[44043093,44043535,44043673,44043729,44044372,44044509,44045072,44045725,44047094,44052299],"created_at":"2025-05-20T15:35:29Z","created_at_i":1747755329,"num_comments":19,"objectID":"44042791","points":33,"story_id":44042791,"story_text":"Hi HN, we're Roya and Michael, co-founders of Opusense AI (https://www.opusense.com/), a tool to help engineers and consultants automatically generate construction site reports from typed or voice notes, plus photos.
Here\u2019s a video: https://www.youtube.com/watch?v=u3Pi1iih1_Y.
Before this, I (Roya) worked in human-machine interaction at Huawei, and before that as a construction site inspector for civil engineering firms. I have a PhD in Civil Engineering, and in my experience reporting was by far the most tedious and mind-numbing part of the job.
You\u2019d walk around a site all day taking short notes (maybe, often you'd rely on memory) and snapping photos, then go to three more sites before finally making it back to the office and try to remember everything you wanted to write. Sometimes you\u2019d fill in gaps from memory or you\u2019d keep it purposefully vague. Reports had to be consistent, branded, and checked by senior engineers. It was a huge time sink across the team.
Writing reports was the worst part of the job, so we built Opusense to get rid of it. On-site, users type or dictate short notes (e.g. \u201crebar exposed east end of slab\u201d), and the tool turns them into full sentences, paragraphs, tables, or photo captions in a report template that matches the firm\u2019s format. You can work offline, and it syncs automatically when back online.
Most inspection and reporting tools are built for checklist-style workflows (which is great for home inspections or punch lists), but civil, structural, environmental, or geotechnical engineers usually need freeform notes, not radio buttons.
This is a particularly good fit for LLMs because engineering field reports live in a constrained, conventional domain: similar language, repeated structures, and highly standardized content across firms and projects. There\u2019s a lot of redundancy and grunt work, summarizing the same site conditions, formatting repetitive data, translating field notes into polished paragraphs, all of which LLMs handle well with the right prompting and guardrails. We\u2019re not generating arbitrary prose; we\u2019re transforming structured inputs (notes, images, forms) into structured outputs, with firm-defined templates and required fields that minimize the risk of hallucination. When facts matter (e.g. test results or measurements), we keep them grounded in the user\u2019s input, the model doesn\u2019t invent data because there\u2019s nothing for it to invent. This makes it one of those cases where LLMs aren\u2019t just a novelty, they're genuinely the best tool for the job.
Under the hood, we use a combination of prompt-engineered LLMs and firm-specific formatting rules to get outputs that don\u2019t just sound good, but also look right. We\u2019ve recently added translation features, and we\u2019re iterating quickly based on field feedback. We charge per seat and are deployed at mid size firms, and trialing with some multinational engineering firms who have thousands of reports to file each week. We're also starting to see interest from construction managers and developers who do their own internal QA reporting.
We don't have a self-serve way to try out the product yet, because the way our business works requires templates to be customized by company. But there\u2019s a demo at https://www.youtube.com/watch?v=u3Pi1iih1_Y, and if you want to poke around the UI yourself, here\u2019s a sample account to log in with:
login: hndemo@opusense.com\n password: OpusenseHacker2025\nWe\u2019d love to hear how others are thinking about tools for field work, reporting, or similar workflows (engineering, architectural, etc.). If you\u2019ve built in this space, or have thoughts on how to improve it, we\u2019re all ears!","title":"Launch HN: Opusense (YC X25) \u2013 AI assistant for construction inspectors on site","updated_at":"2025-10-06T20:22:56Z"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"jillesvangurp"},"comment_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["which","ai","model","is","best","for","summarizing"],"value":"> How does AI attend meetings, ask questions of local officials or bodies, provide in-person reports and original fact based copy?
Is that a thing with local journalism now? Actual journalists still doing stuff like that? Lets not pretend that a lot of good writing or journalism is involved with local news papers. It's not a particularly hard job for an LLM to do better.
In any case to answer your questions:
- yes you can let an AI listen in on meetings and conversations and they'll do a decent enough job of transcribing what was said and summarizing key points.
- you can also use them to prepare for interviews, get some background information on people, companies, etc. and prepare some questions. Perplexity is great for background research like that.
- and if you provide all of that as context, it will do a half decent job of writing an article. You might want to do some prompt engineering to tune style and content and maybe fact check a thing or two. For best results, maybe use one model to generate and another to fact check.
You are probably already reading a lot of stuff that is AI generated; or at least AI assisted without realizing it. It's not always that obvious. The future is last year. Already happened. What you are seeing is the more obvious/lazy stuff. Which is of course a thing because good enough is good enough and the benchmark for that is very low with most media corporations. And being factual isn't necessary something they value either."},"story_title":{"fullyHighlighted":false,"matchLevel":"partial","matchedWords":["ai","is"],"value":"AI slop is already invading Oregon's local journalism"},"story_url":{"matchLevel":"none","matchedWords":[],"value":"https://www.opb.org/article/2024/12/09/artificial-intelligence-local-news-oregon-ashland/"}},"_tags":["comment","author_jillesvangurp","story_42378673"],"author":"jillesvangurp","comment_text":"> How does AI attend meetings, ask questions of local officials or bodies, provide in-person reports and original fact based copy?
Is that a thing with local journalism now? Actual journalists still doing stuff like that? Lets not pretend that a lot of good writing or journalism is involved with local news papers. It's not a particularly hard job for an LLM to do better.
In any case to answer your questions:
- yes you can let an AI listen in on meetings and conversations and they'll do a decent enough job of transcribing what was said and summarizing key points.
- you can also use them to prepare for interviews, get some background information on people, companies, etc. and prepare some questions. Perplexity is great for background research like that.
- and if you provide all of that as context, it will do a half decent job of writing an article. You might want to do some prompt engineering to tune style and content and maybe fact check a thing or two. For best results, maybe use one model to generate and another to fact check.
You are probably already reading a lot of stuff that is AI generated; or at least AI assisted without realizing it. It's not always that obvious. The future is last year. Already happened. What you are seeing is the more obvious/lazy stuff. Which is of course a thing because good enough is good enough and the benchmark for that is very low with most media corporations. And being factual isn't necessary something they value either.","created_at":"2024-12-10T22:48:50Z","created_at_i":1733870930,"objectID":"42382547","parent_id":42379425,"story_id":42378673,"story_title":"AI slop is already invading Oregon's local journalism","story_url":"https://www.opb.org/article/2024/12/09/artificial-intelligence-local-news-oregon-ashland/","updated_at":"2024-12-11T01:54:39Z"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"altairprime"},"comment_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["which","ai","model","is","best","for","summarizing"],"value":"Hmm. Well, I didn't see any particular weaknesses in the evidence, but I'm already assuming all discussion of this topic is forward-looking conjecture by third parties with incomplete knowledge until either Nvidia internal data leaks, or they get penalized by the SEC. Still, scanning the article, here are key phrases that should have indicated 'this is an opinion' to readers:
> Here is my take
'take' is a fancy word for 'opinionated interpretation', implying opinion.
> even just connecting the dots myself (with the help of Gemini)
If an AI was involved in producing this writing, it could be whole-cloth fiction, so I certainly would not attribute factualness to anything in the entire post that I didn't have an independent source for. Others would differ, but even setting aside this particular point, there are so many more.
> it feels like their biggest customers
This does not particularly give confidence that they're stating facts. If they were stating facts, they would say simply "their biggest customers are" without the weakening-disclaimer language of "it feels like" or "I suspect that" or etc.
> My personal read?
This is explicitly a rhetorical device indicating personal interpretation of the three pieces of data listed above, cited from their published financials. Obviously one should double-check the financials to confirm that "Gemini" didn't make shit up, but either these three bulleted-list items are factually erroneous, debatable interpretations, or factually correct. This can be specifically addressed if desired.
> I didn't discover this next part
The author is summarizing someone else's work here. I've read other authors on the same topic as well. This is not, as I would say, 'primary source' material, and if their interpretation is bogus, it's on me for relying on it (cc Gemini involvement) rather than tracking down the original sources (which I did, earlier).
> they look more like
This is a normal personal interpretation signifier.
> my guess is
This is an explicit theorising signifier.
And so, having done that exercise and read through the entire post, I fail to identify claims made by the author that are "presented with the tone and certainty of established fact". The author presents zero facts, as far as I can tell, in a plain reading. Am I missing some specific instance where they make a factual claim that isn't unambiguously weakened by the repeated contextual 'this is opinion, this is interpretation, referring to the work of others, published financials' clues that are present throughout, regardless of how one interprets the stated use of Gemini?
This is, I think, at the core of where I'm confused about your opinions here today. You've stated opinions about the work \u2014 and yes, even a neutral summary is opinionated! \u2014 but even this far deep in the discussion, you still haven't referred to actual segments of the actual work to explain how you reached your opinion. When my reply is confusion \u2014 i.e., "I don't follow, could you refer to specific quotes from the post?" to each of your objections: "when speculation is presented", "it states mechanisms and outcomes", "blurring the line between assumption and demonstration", and "glossing over the range of alternative explanations" \u2014 then I certainly empathize with others who refuse to respond. I've tried my very best to give you the benefit of doubt, but I'm just lost at this point; your opinion is unsupported general statements with zero specifics, and armchair dentistry is not most people's idea of fun when it comes to getting someone to explain how they formed an opinion (especially in today's world where "an AI generated these confusingly-general statements" is a high-probability outcome).
Perhaps an example will help convey the confusion.
> The issue is when speculation is presented with the tone and certainty of established fact. The article doesn\u2019t merely offer possibilities in light of missing data; it states mechanisms and outcomes as though the evidence for them is already in hand. So the objection isn\u2019t to building a model, but to blurring the line between assumption and demonstration, and to glossing over the range of alternative explanations that the same incomplete information could support.
Regarding one of your claims in your comment, I agree with your interpretation of the post itself, but I do not share your negative opinion of that interpretation.
My above paragraph is completely serious. I mean every word of it, and it's not a constructed example. It's an actual response I had to your post. How would you respond? Most people would ask, 'Which claim led you to that interpretation?', and that's precisely the question that I'm left with for each of your interpretations here today. That's why I chose it as an example: it's a content-free opinion, that carries only judgment but not content or meaning. Without the essential core of what I'm disagreeing with, what use is it to you that I disagree at all? What useful contribution have I made?
I hope this will help set you on a more productive course with the HN community for future posts."},"story_title":{"matchLevel":"none","matchedWords":[],"value":"NVIDIA frenemy relation with OpenAI and Oracle"},"story_url":{"matchLevel":"none","matchedWords":[],"value":"https://philippeoger.com/pages/deep-dive-into-nvidias-virtuous-cycle"}},"_tags":["comment","author_altairprime","story_46196076"],"author":"altairprime","comment_text":"Hmm. Well, I didn't see any particular weaknesses in the evidence, but I'm already assuming all discussion of this topic is forward-looking conjecture by third parties with incomplete knowledge until either Nvidia internal data leaks, or they get penalized by the SEC. Still, scanning the article, here are key phrases that should have indicated 'this is an opinion' to readers:
> Here is my take
'take' is a fancy word for 'opinionated interpretation', implying opinion.
> even just connecting the dots myself (with the help of Gemini)
If an AI was involved in producing this writing, it could be whole-cloth fiction, so I certainly would not attribute factualness to anything in the entire post that I didn't have an independent source for. Others would differ, but even setting aside this particular point, there are so many more.
> it feels like their biggest customers
This does not particularly give confidence that they're stating facts. If they were stating facts, they would say simply "their biggest customers are" without the weakening-disclaimer language of "it feels like" or "I suspect that" or etc.
> My personal read?
This is explicitly a rhetorical device indicating personal interpretation of the three pieces of data listed above, cited from their published financials. Obviously one should double-check the financials to confirm that "Gemini" didn't make shit up, but either these three bulleted-list items are factually erroneous, debatable interpretations, or factually correct. This can be specifically addressed if desired.
> I didn't discover this next part
The author is summarizing someone else's work here. I've read other authors on the same topic as well. This is not, as I would say, 'primary source' material, and if their interpretation is bogus, it's on me for relying on it (cc Gemini involvement) rather than tracking down the original sources (which I did, earlier).
> they look more like
This is a normal personal interpretation signifier.
> my guess is
This is an explicit theorising signifier.
And so, having done that exercise and read through the entire post, I fail to identify claims made by the author that are "presented with the tone and certainty of established fact". The author presents zero facts, as far as I can tell, in a plain reading. Am I missing some specific instance where they make a factual claim that isn't unambiguously weakened by the repeated contextual 'this is opinion, this is interpretation, referring to the work of others, published financials' clues that are present throughout, regardless of how one interprets the stated use of Gemini?
This is, I think, at the core of where I'm confused about your opinions here today. You've stated opinions about the work \u2014 and yes, even a neutral summary is opinionated! \u2014 but even this far deep in the discussion, you still haven't referred to actual segments of the actual work to explain how you reached your opinion. When my reply is confusion \u2014 i.e., "I don't follow, could you refer to specific quotes from the post?" to each of your objections: "when speculation is presented", "it states mechanisms and outcomes", "blurring the line between assumption and demonstration", and "glossing over the range of alternative explanations" \u2014 then I certainly empathize with others who refuse to respond. I've tried my very best to give you the benefit of doubt, but I'm just lost at this point; your opinion is unsupported general statements with zero specifics, and armchair dentistry is not most people's idea of fun when it comes to getting someone to explain how they formed an opinion (especially in today's world where "an AI generated these confusingly-general statements" is a high-probability outcome).
Perhaps an example will help convey the confusion.
> The issue is when speculation is presented with the tone and certainty of established fact. The article doesn\u2019t merely offer possibilities in light of missing data; it states mechanisms and outcomes as though the evidence for them is already in hand. So the objection isn\u2019t to building a model, but to blurring the line between assumption and demonstration, and to glossing over the range of alternative explanations that the same incomplete information could support.
Regarding one of your claims in your comment, I agree with your interpretation of the post itself, but I do not share your negative opinion of that interpretation.
My above paragraph is completely serious. I mean every word of it, and it's not a constructed example. It's an actual response I had to your post. How would you respond? Most people would ask, 'Which claim led you to that interpretation?', and that's precisely the question that I'm left with for each of your interpretations here today. That's why I chose it as an example: it's a content-free opinion, that carries only judgment but not content or meaning. Without the essential core of what I'm disagreeing with, what use is it to you that I disagree at all? What useful contribution have I made?
I hope this will help set you on a more productive course with the HN community for future posts.","created_at":"2025-12-09T02:50:35Z","created_at_i":1765248635,"objectID":"46200708","parent_id":46197268,"story_id":46196076,"story_title":"NVIDIA frenemy relation with OpenAI and Oracle","story_url":"https://philippeoger.com/pages/deep-dive-into-nvidias-virtuous-cycle","updated_at":"2026-03-05T23:06:36Z"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"PaulHoule"},"comment_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["which","ai","model","is","best","for","summarizing"],"value":"This book talks about that problem
https://en.wikipedia.org/wiki/G%C3%B6del,_Escher,_Bach
and makes it clear the "truth" is not so simple as to be a module added to the LLM.
Logical and mathematical reasoning is rather specialized, but it is a useful feature. Understanding text, though, frequently involves setting up and solving logic problems and solving it to be sure your interpretation of the arguments is correct, all the more true if you expect it to read about a system and then apply that system to another text. So you at least run into the NP-complete world of SMT solving and the system must realize the requirements of 1980s symbolic AI no matter what technology is under the hood.
It's much worse than that because it is reasoning with uncertainty. If I were uncharitable I'd say ChatGPT was wrong if it said the universe is composed like this
https://en.wikipedia.org/wiki/Lambda-CDM_model
and a year later we heard otherwise. At a very high level it had best be able to explain that there are alternate views on the subject, but it has to know when to stop. If I was asking "How to get the length of a string in Python?" is not helpful to "teach the controversy" that some of us use "sum(1 for c in s)". It has to handle contradictions such as person A and person B believing different things, the same person believing different things at different times, plus problems that are impossible to solve or practically impossible so logic is further complicated.
One route links the LLM up with an "old AI" system like the way AlphaGo links up neural and search based players.
The straight route for LLMs is to enlarge the attention capacity. Right now ChatGPT has a sub-words 4096 token attention window. If the task is "write two pages summarizing topic T with citations to the literature" it has to read all of the papers it cites. That could be 400,000 to 4,000,000 tokens which one could presenting to the LLM which would be 100-1000x bigger. Maybe it can swap inputs in and out and otherwise conserve space, but I think the big weakness of LLM in practice is that they run out of steam if the text is larger than what they are built for."},"story_title":{"fullyHighlighted":false,"matchLevel":"partial","matchedWords":["ai"],"value":"Ask HN: How long until AI shows general intelligence without failures?"}},"_tags":["comment","author_PaulHoule","story_34661263"],"author":"PaulHoule","comment_text":"This book talks about that problem
https://en.wikipedia.org/wiki/G%C3%B6del,_Escher,_Bach
and makes it clear the "truth" is not so simple as to be a module added to the LLM.
Logical and mathematical reasoning is rather specialized, but it is a useful feature. Understanding text, though, frequently involves setting up and solving logic problems and solving it to be sure your interpretation of the arguments is correct, all the more true if you expect it to read about a system and then apply that system to another text. So you at least run into the NP-complete world of SMT solving and the system must realize the requirements of 1980s symbolic AI no matter what technology is under the hood.
It's much worse than that because it is reasoning with uncertainty. If I were uncharitable I'd say ChatGPT was wrong if it said the universe is composed like this
https://en.wikipedia.org/wiki/Lambda-CDM_model
and a year later we heard otherwise. At a very high level it had best be able to explain that there are alternate views on the subject, but it has to know when to stop. If I was asking "How to get the length of a string in Python?" is not helpful to "teach the controversy" that some of us use "sum(1 for c in s)". It has to handle contradictions such as person A and person B believing different things, the same person believing different things at different times, plus problems that are impossible to solve or practically impossible so logic is further complicated.
One route links the LLM up with an "old AI" system like the way AlphaGo links up neural and search based players.
The straight route for LLMs is to enlarge the attention capacity. Right now ChatGPT has a sub-words 4096 token attention window. If the task is "write two pages summarizing topic T with citations to the literature" it has to read all of the papers it cites. That could be 400,000 to 4,000,000 tokens which one could presenting to the LLM which would be 100-1000x bigger. Maybe it can swap inputs in and out and otherwise conserve space, but I think the big weakness of LLM in practice is that they run out of steam if the text is larger than what they are built for.","created_at":"2023-02-05T16:47:55Z","created_at_i":1675615675,"objectID":"34666369","parent_id":34661263,"story_id":34661263,"story_title":"Ask HN: How long until AI shows general intelligence without failures?","updated_at":"2024-09-20T13:16:56Z"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"YeGoblynQueenne"},"comment_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["which","ai","model","is","best","for","summarizing"],"value":"The title of the article does an awful job of summarising it. The main subject of the article is not about the level of intelligence of current or future AI, as the title suggsts. Instead, the article is a reflection on the progress in the field in the last few years and a discussion of the degree to which progress in deep learning has benefited, or harmed, AI research in general.
This is best summarised by the "Key Insights" box on top of the article:
> the recent successes of deep learning have revealed something very interesting about the structure of our world, yet this seems to be the least pursued and talked about topic today
> In AI, the key question today is not whether we should use model-based or function-based approaches but how to integrate and fuse them so we can realise their objective benefits
> We need a new generation of AI researchers who are well versed in an appreciate classical AI, machine learning, and computer science more broadly while also being informed about AI history.
The first "key point" refers to classes of functions that can be seen as "cognitive functions". For example, mapping a set of inputs to outputs can reasonably be considered as approximating some aspect of cognition when the inputs are regions of images and the subjects their labels, so that the function performs object recognition, a task that AI research has long considered an aspect of cognition. Understanding how such functions work has the potential to contribute to our understanding of human cognition, that has long been a major goal of AI research. Yet, in recent years, interest has shifted from understanding such results to applying them at the level of phone apps, etc.
The final key point is a call to arms. We can't make progress as a field by throwing out everything we've done before, everytime we achieve some success in a narrow range of tasks. The author witnessed this happenning in the 1980's with the rise and fall of expert systems -and the AI winter that followed. In modern times, the success of deep learning has all but eclipsed the deep knowledge that researchers in the field once possessed about symbolic logic and important avenues of research are impossible to follow because the younger generation of researchers simply don't have the necessary background - and are "bullied by the success" of neural networks into directing their careers towards neural network research, whatever their true interests.
On a personal level, not summarising the article anymore, the latter is the most disturbing development. Neural networks can perform "perceptual" tasks, but are wholly incapable of reasoning. Symbolic AI had reasoning down pat- and not in approximate fashion (as a recent trend in deep learning research attempts to perform it). Yet, we seem to have regressed and lost one ability to perform one set of cognitive tasks in the process of figuring out how to perform another.
In the past, AI researchers were well-rounded polymaths, versed in CS but also (continuous) mathematics, physics, psychology, linguistics... Nowadays, researchers seem to be optimising for a narrow band of knowledge and ignoring everything else. This cannot end well."},"story_title":{"matchLevel":"none","matchedWords":[],"value":"Human-Level Intelligence or Animal-Like Abilities? (2018)"},"story_url":{"matchLevel":"none","matchedWords":[],"value":"https://cacm.acm.org/magazines/2018/10/231373-human-level-intelligence-or-animal-like-abilities/fulltext"}},"_tags":["comment","author_YeGoblynQueenne","story_23438359"],"author":"YeGoblynQueenne","comment_text":"The title of the article does an awful job of summarising it. The main subject of the article is not about the level of intelligence of current or future AI, as the title suggsts. Instead, the article is a reflection on the progress in the field in the last few years and a discussion of the degree to which progress in deep learning has benefited, or harmed, AI research in general.
This is best summarised by the "Key Insights" box on top of the article:
> the recent successes of deep learning have revealed something very interesting about the structure of our world, yet this seems to be the least pursued and talked about topic today
> In AI, the key question today is not whether we should use model-based or function-based approaches but how to integrate and fuse them so we can realise their objective benefits
> We need a new generation of AI researchers who are well versed in an appreciate classical AI, machine learning, and computer science more broadly while also being informed about AI history.
The first "key point" refers to classes of functions that can be seen as "cognitive functions". For example, mapping a set of inputs to outputs can reasonably be considered as approximating some aspect of cognition when the inputs are regions of images and the subjects their labels, so that the function performs object recognition, a task that AI research has long considered an aspect of cognition. Understanding how such functions work has the potential to contribute to our understanding of human cognition, that has long been a major goal of AI research. Yet, in recent years, interest has shifted from understanding such results to applying them at the level of phone apps, etc.
The final key point is a call to arms. We can't make progress as a field by throwing out everything we've done before, everytime we achieve some success in a narrow range of tasks. The author witnessed this happenning in the 1980's with the rise and fall of expert systems -and the AI winter that followed. In modern times, the success of deep learning has all but eclipsed the deep knowledge that researchers in the field once possessed about symbolic logic and important avenues of research are impossible to follow because the younger generation of researchers simply don't have the necessary background - and are "bullied by the success" of neural networks into directing their careers towards neural network research, whatever their true interests.
On a personal level, not summarising the article anymore, the latter is the most disturbing development. Neural networks can perform "perceptual" tasks, but are wholly incapable of reasoning. Symbolic AI had reasoning down pat- and not in approximate fashion (as a recent trend in deep learning research attempts to perform it). Yet, we seem to have regressed and lost one ability to perform one set of cognitive tasks in the process of figuring out how to perform another.
In the past, AI researchers were well-rounded polymaths, versed in CS but also (continuous) mathematics, physics, psychology, linguistics... Nowadays, researchers seem to be optimising for a narrow band of knowledge and ignoring everything else. This cannot end well.","created_at":"2020-06-08T20:58:41Z","created_at_i":1591649921,"objectID":"23460739","parent_id":23438359,"story_id":23438359,"story_title":"Human-Level Intelligence or Animal-Like Abilities? (2018)","story_url":"https://cacm.acm.org/magazines/2018/10/231373-human-level-intelligence-or-animal-like-abilities/fulltext","updated_at":"2024-09-20T06:24:39Z"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"lhl"},"comment_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["which","ai","model","is","best","for","summarizing"],"value":"We do live in an age of frontier LLMs... For fun, I'll just use Kimi K2 (on Kagi Assistant).
> Can you explain what this means and its significance? Assume that I'm a layperson with no familiarity with LLM jargon so explain all of the technical terms, references, names. https://github.com/MoonshotAI/Kimi-Linear
Imagine your brain could only \u201clook at\u201d a few words at a time when you read a long letter. Today\u2019s big language models (the AI that powers chatbots) have the same problem: the longer the letter gets, the more scratch paper they need to keep track of it all. That scratch paper is called the \u201cKV cache,\u201d and for a 1 000 000-word letter it can fill a small library.
Kimi Linear is a new way for the AI to read and write that throws away most of that scratch paper yet still understands the letter. It does this by replacing the usual \u201clook at every word every time\u201d trick (full attention) with a clever shortcut called linear attention. The shortcut is packaged into something they call Kimi Delta Attention (KDA).
What the numbers mean in plain English
51.0 on MMLU-Pro: on a 4 000-word school-test set, the shortcut scores about as well as the old, slow method.\n 84.3 on RULER at 128 000 words: on a much longer test it keeps the quality high while running almost four times faster.\n 6 \u00d7 faster TPOT: when the AI is writing its reply, each new word appears up to six times sooner than with the previous best shortcut (MLA).\n 75 % smaller KV cache: the scratch paper is only one-quarter the usual size, so you can fit longer conversations in the same memory.\n Full attention: the old, accurate but slow \u201clook back at every word\u201d method.\n KV cache: the scratch paper that stores which words were already seen.\n Linear attention: a faster but traditionally weaker way of summarising what was read.\n Gated DeltaNet: an improved linear attention trick that keeps the most useful bits of the summary.\n Kimi Delta Attention (KDA): Moonshot\u2019s even better version of Gated DeltaNet.\n Hybrid 3:1 mix: three layers use the fast KDA shortcut, one layer still uses the old reliable full attention, giving speed without losing smarts.\n 48 B total, 3 B active: the model has 48 billion total parameters but only 3 billion \u201cturn on\u201d for any given word, saving compute.\n Context length 1 M: it can keep track of about 1 000 000 words in one go\u2014longer than most novels.\n> Can you explain what this means and its significance? Assume that I'm a layperson with no familiarity with LLM jargon so explain all of the technical terms, references, names. https://github.com/MoonshotAI/Kimi-Linear
Imagine your brain could only \u201clook at\u201d a few words at a time when you read a long letter. Today\u2019s big language models (the AI that powers chatbots) have the same problem: the longer the letter gets, the more scratch paper they need to keep track of it all. That scratch paper is called the \u201cKV cache,\u201d and for a 1 000 000-word letter it can fill a small library.
Kimi Linear is a new way for the AI to read and write that throws away most of that scratch paper yet still understands the letter. It does this by replacing the usual \u201clook at every word every time\u201d trick (full attention) with a clever shortcut called linear attention. The shortcut is packaged into something they call Kimi Delta Attention (KDA).
What the numbers mean in plain English
51.0 on MMLU-Pro: on a 4 000-word school-test set, the shortcut scores about as well as the old, slow method.\n 84.3 on RULER at 128 000 words: on a much longer test it keeps the quality high while running almost four times faster.\n 6 \u00d7 faster TPOT: when the AI is writing its reply, each new word appears up to six times sooner than with the previous best shortcut (MLA).\n 75 % smaller KV cache: the scratch paper is only one-quarter the usual size, so you can fit longer conversations in the same memory.\n Full attention: the old, accurate but slow \u201clook back at every word\u201d method.\n KV cache: the scratch paper that stores which words were already seen.\n Linear attention: a faster but traditionally weaker way of summarising what was read.\n Gated DeltaNet: an improved linear attention trick that keeps the most useful bits of the summary.\n Kimi Delta Attention (KDA): Moonshot\u2019s even better version of Gated DeltaNet.\n Hybrid 3:1 mix: three layers use the fast KDA shortcut, one layer still uses the old reliable full attention, giving speed without losing smarts.\n 48 B total, 3 B active: the model has 48 billion total parameters but only 3 billion \u201cturn on\u201d for any given word, saving compute.\n Context length 1 M: it can keep track of about 1 000 000 words in one go\u2014longer than most novels.\n