Not quite - as I understand it box-counting measures global space-filling, manifolds handle local coordinate structure. Consider that the Earth is locally flat but globally spherical, and a Möbius strip vs cylinder are locally identical but globally different. Related problems, but the tools reveal different aspects of geometry. So I think whether “this is exactly what topological manifolds are for” depends what you’re trying to understand.
Synthi is an open web tool that instantly summarizes and synthesizes Hacker News threads and their linked articles, grouping every point of view by topic.
I love the deep discussions on HN, but I never have time to read a long article and a 400-comment thread. The tipping point for building this was when I nearly burned through my monthly API credits on another service just trying to synthesize a few threads! I needed my own tool.
How It Works:
1. Paste any HN thread URL into Synthi.
2. It instantly detects it and fetches the linked article.
3. Click "Full Analysis" for a unified, topic-based synthesis (with attribution to the article or the commenter).
4. Export the result, bookmark it, or listen to it (the output is text-to-speech friendly).
Key Features:
* Smart HN workflow: Auto-detects HN links for a one-click analysis.
* Works with any URLs: Can also synthesize any two articles or pieces of text.
* 100% client-side: All processing happens in your browser. No backend, no tracking.
* Open source (MIT License): The code is yours to inspect, fork, and use.
Synthi is a "Bring Your Own Key" app. You'll need a Google Gemini API key, which is stored securely in your browser's local storage and never sent anywhere else.
For now, it's Gemini only, largely because the free tier on Google AI Studio is incredibly generous and a great way to get started. I'm considering adding support for other models like Claude or OpenAI via OpenRouter in the future.
I've been using this every day and I hope it's useful to some of you too. All feedback is welcome!
Wow - I'd forgotten all about this but just realized I have posts from an entire phase of earlier professional life - topic by topic and event by event - on an old blog there. Amazingly the browser remembered my login so I was able to find the URL. It's been quite a trip down memory lane revisiting some of the posts. Not sure I need to keep any of that published but I'll at least scrape and store it somewhere for old times sake. Maybe I'll find some buried gem of an idea when I scan them during the great scrape. Or - optimistically - perhaps a future zillion-token context LLM will uncover some personal patterns that unleash deep and actionable insights. Irrespective of the measurable value, I just hate to see the old posts dissapear forever.
In 1977 you said that computers were answers in search of questions. Has that changed?
Well, the types of computers we have today are tools. They’re responders: you ask a computer to do something and it will do it. The next stage is going to be computers as “agents.” In other words, it will be as if there’s a little person inside that box who starts to anticipate what you want. Rather than help you, it will start to guide you through large amounts of information. It will almost be like you have a little friend inside that box. I think the computer as an agent will start to mature in the late '80s, early '90s…
You’d start to teach it about yourself. And it would just keep storing all this information about you and maybe it would recognize that every Friday afternoon you like to do something special, and maybe you’d like it to help you with this routine. So about the third time it asks you: “Well, would you like me to do this for you every Friday?” You say, “Yes,” and before long it becomes an incredibly powerful helper. It goes with you everywhere you go. It knows most of the raw information in your life that you’d like to keep, but then starts to make connections between things, and one day when you’re 18 and you’ve just split up with your girlfriend it says: “You know, Steve, the same thing has happened three times in a row.”
This is a fun project to be sure. I just wish the author would not refer to the experiment as an "autonomous startup builder" unless they mean it humorously. Having poked around the GitHub repo and read through the materials, it seems like more of an AI coding assistant running in a loop that built and deployed a broken web application with no users, no business model, and no understanding of what problem it was trying to solve. There were quasi-autonomous processes and there were things that were built, but nothing I'd call a startup.
One of the things that I found most frustrating about USB-C hubs is how hard it is to find one that actually gives you multiple USB-C ports. I have several USB-C devices but most hubs just give you one USB-C port and a bunch of USB-A ports. At most it’s 2 USB-C ports but only with the hub that plugs into both USB-C ports on my MacBook Pro (so I’m never able to get more ports than I started with). The result is I end up having to keep swapping devices. For a connector that was supposed to be the "one universal port," it's weird that most hubs assume you only need one USB-C connection. Has anyone found a decent hub with multiple USB-C data outputs?
I'm in the same boat. It seems like the mindset for consumer-grade hubs is to provide support for as many old, legacy devices as possible, rather than a higher number of new devices.
Another problem is that USB-A ports are dirt cheap and simple to implement, so hub makers feel like "leaving free IO on the table" by not sprinkling them on everything. Whereas each "decent" USB-C port has enough complexity to think twice about adding it.
Nevertheless, there are a couple of options. Try searching for "USB-C only hub". You will get some results, but they are basically the identical product (same IO card), just with different housings. So you can pretty much count with these specs: 1 USB-C in for power, 3–4 USB-C out, 5 or 10Gbps each, Power Delivery at various wattages. No video support.
I have one of these on my desk right now, it's from the brand "Minisopuru", I get power and four USB-C "3.2 Gen 2" ports. It's fine. But like I said, it's no Thunderbolt, and no video support, so I have to "waste" the other port on my MacBook just for my external display.
There are also Thunderbolt / USB4 devices which will give you a bonkers amount of IO, including good TB / USB-C ports usually (plus some USB-A of course, as a spit in the face – so you'd need to ignore those). But these are not hubs, they are docks, which is a different product class entirely (big and heavy, more expensive, dedicated power supply).
Something I've been doing recently to salvage the USB-A ports I still begrudgingly encounter, while continuing to (force myself to) upgrade all my devices to type-C, are these: [0]. 1-to-1 USB-A male to USB-C female adapters. I just stick them in all USB-A ports I encounter, leave them there all the time, and move on with my life. It's a bit bulky and looks kinda stupid, but it basically gives me USB-C everywhere I need (including work-issued PCs and docking stations) for just a couple of bucks. For low-bandwidth devices like headphones, keyboard / mice / gamepads, or even my phone, it works perfectly fine.
You can get them now. Thunderbolt and USB 4 hubs will often have multiple USB C ports and only need one plug. I have one that's more of a docking station:
> One of the things that I found most frustrating about USB-C hubs is how hard it is to find one that actually gives you multiple USB-C ports.
It's the power consumption.
IIRC, USB-C has a base power per port of 15W (5V @ 3A) with just basic CC resistors. USB 2 starts at 0.5W (5V @ 0.1A) and is only supposed to allow 2.5W (5V @ 0.5A) after negotiation. USB 3 is 4.5W (5V @ .900A).
Note that the Caldigit hub linked in a sibling has a power supply of 20V @ 9A. That's 180W!
Yes, I've bought a chinese ("Acasis" brand) TB4 hub which has three TB4 downstream ports and an USB 3.x hub with three downstream 10 Gbps USB-C ports. There are also weird combos like one downstream TB3 + three downstream USB-C 3.x. Still not great, but it's better than a single downstream port.
I’m still stuck on the first sentence "An LLM should never output anything but tool calls and their arguments” because it just doesn’t make sense to me.
Tool calling is great, but LLMs are - and should be used as - more than just tool callers. I mean, some tools will have to be other LLMs doing what they’re good at, like writing a novel, summarizing, brainstorming ideas, or explaining complex topics. Tools are useful, but the stuff LLMs actually do is also useful. The basic premise that LLMs should never output anything beyond tools and arguments is leaving most of the value of LLMs on the table.
I think the blog simply does not explain well. Consider the example of a text editor, the "tool calls" are text fragments generated by the LLM then embedded into text editor tool calls that place the generated text fragment into the text editor, performing cuts, pastes, and so on.
FWIW, I've done this and it works incredibly well. It's essentially integrating the LLM into the text editor, and requests of the LLM are more like requests of the text editor directly. The mental model I use is the editor has become an AI Agent itself. I've also done with with spreadsheets, web page editors, various tools in project management software. It's an incredible perspective that works.
Got it, thanks for clarifying! So if I’m understanding you right, you’re saying that all the generative stuff the LLM does—like creating text—basically becomes part of the ‘arguments’ the original post talks about, and then that gets paired with a tool call (like inserting into a text editor, doing edits, etc.). I was focused on the tool call not the argument content aspect of the post.
And it sounds like you’ve had a lot of success with this approach in an impressive variety of application types. May I ask what tooling you usually use for this (eg custom python for each hack? MCP? some agent framework like LangGraph/ADK/etc, other?)
I noticed fairly early that the foundation LLMs have the source code to most FOSS, as well as the developer conversations, the user discussions trying to understand how to use that software, and the documentation too. The foundational models have a good amount of training data of each popular FOSS app, and by examining the code and the developer comments, and then adopting their language style, the LLM practically takes on the persona of the developer. So I spent some time understanding the internal communications of each app, and my 'tool calls' are structured JSON of the internal structures these applications use, and my own code receives these structured outputs and I just replace in the application's running memory. Not quite so blind as I describe, some of the insertion of these data structures is complicated.
In the end, each app is both what it was before, as well as can be driven by prompts. I've also specialized each to have 4 agents that are as I describe, but they each have a different representation of the app's internal data; for example, a word processor has the "content, the document" in HTML/CSS as well as raw text. When one wants to manipulate the text, requests use the HTML/CSS representation, and selections go through a slightly separate logic than a request to be applied to the entire document. When one wants to critically analyze the text, it is ASCII text, no need for the HTML/CSS at all. When one wants to use the document as a knowledge base, outside the editor, that's yet another variant that uses the editor to output a RAG ready representation.
That system would make a tidy startup, especially if tightly integrated with an open source office suite behind the scenes (LibreOffice, OpenOffice, etc) and a generative AI native UX.
> No one actually pays that price. The $1000 misrepresented in the article…
With respect, that is absolutely incorrect. People absolutely pay over $1000 and do so monthly. For example, Kaiser of Northern California makes it very difficult for their doctors to prescribe these, and nearly impossible to get a prescription for Monjaro (which is particularly effective). Therefore, Kaiser patients/insured for whom these drugs are of immense benefit but who must have their prescriptions from out of network physicians receive ZERO insurance coverage. This means they get neither the negotiated insurance price discount nor any co-pay on the full cost. I am directly aware of this. And it is a travesty. Yet the benefits of these drugs is so significant and uniquely available through these drugs that in a sense, if it is possible to pay, then pay one must. Because in effect they are invaluable.
They would buy direct from LLY for $550/mo or pay the $650/mo via manufacturer coupon for Zepbound. If their primary care provider refuses to prescribe telehealth will trivially do it in 10 minutes for about $150 every 6mo.
The only folks not covered by the latter mostly either have no insurance or Medicare.
That said, many folks take a few months to figure out these programs exist, and some consumers simply never do the research.
I was fortunate to get early access to the new Agent SDK and APIs that OpenAI dropped today and made an open source project to show some of the capabilities [1]. If you are using any of the other agent frameworks like LangGraph/LangChain, AutoGen, Crew, etc I definitely suggest giving this agent SDK a spin.
To ease into it, I added the entire SDK with examples and full documentation as a single text file in my repo [2] so you can quickly get up to speed be adding it to a prompt and just asking about it or getting some quick start code to play around with.
The code in my repo is very modular so you can try implementing any module using one of the other frameworks to do a head-to-head.
Here’s a blog post with some more thoughts on this SDK [3] and some if its major capabilities.
I’ve been putting Grok 3 through some challenging test flows, and it’s impressing me. Not flawless, but the quality is generally exceptional, and the speed? Lightning-fast.
It’s delivering answers on par with the other top models on my back-pocket tests and in a fraction of the time. The integrated search is solid overall, though it occasionally misses key resources—likely a web indexing quirk. That said, its ability to ingest content from any website when you provide the URL is a standout feature.
The analysis is razor-sharp. It’s picking up on nuances—and even uncovering entirely new angles—that other models have overlooked. I just posed a tough hypothetical about a new line of business I’m looking at for my consulting service and it identified an entirely new sphere of possibility that I hadn’t seen nor had any of the other top models (not Gemini 2.0 Pro, OpenAI’s o1-Pro and o3-mini-high, or Claude 3.5 Sonnet).
An API will be needed for real programmatic testing and certainly for integrated business applications and workflows, but I’m told one is coming in the next few weeks.
It’s still early (we’re talking hours here), but so far, I’m digging it.
It’s gracious of you to say that you’d be sorry, and I did run my comment through 4o (perhaps ironically) which caught a slew of typos and weird grammar issues and offered some improvements. But the robotic sound and anything else you don’t like are my own responsibility. Do you, perhaps, have any thoughts on the substance of the comment?
That's discomforting. My practice of sprinkling em-dashes like salt on a salad dates from my early days on various video game communities' forums. They comfortably mimic interrupted speech in writing. I hope I won't have to soon defend myself against accusations of AI usage just because I belong to the minority that read The Punctuation Guide[0] or a related resource.
It's really the em dash along with superfluous language. I suspect you are fine. Models like 4o have a very specific pattern when folks don't specify their writing style.
- Very 'forced' expressions (back-pocket tests, 'The analysis is razor-sharp')
- The fact you're glazing AI so much means you probably uses it, it's like how it was with crypto bros during all the web3 stuff
- Lack of any substance, like, what does that post say? It regurgitates praises over the AI, but the only tangible feature you mention is the fact it can receive an URL as it's input
Hmmmm it is hard to really place the issue. I am very much in the bullish on AI camp but I don't like writing for the sake of writing and some of the models (4o in this case) have very obvious tells and write in such a way that it takes away from what substance may exist.
One thing that concerns me is when you can't tell whether the comment was authored or just edited by AI. I'm uncomfortable with the idea that HN threads and reddit comments gradually tend towards the grey generic writing style of LLMs, but I don't really mind (save for the prospect of people not learning things they might otherwise!) when comments are edited (i.e. minor changes) for the sake of cleanliness or fixing issues.
I just re-read the post twice and I couldn't find any of the points you mentioned (again, other than using URLs in the input):
- Informal Benchmarks: I'm sorry, what? He mentions 'It’s picking up on nuances—and even uncovering entirely new angles—that other models have overlooked' and 'identified an entirely new sphere of possibility that I hadn’t seen nor had any of the other top models'. Not only it is complete horseshit by itself, but it does not benchmark in any way or form against the mentioned competitors. It's the exact stuff I'd expect out of a LLM.
- Real-World Test Case: As mentioned above, complete horseshit.
- 2 Concrete Features: Yes, I mentioned URLs in the input. I didn't consider 'Integrated Search' (which I'm assuming is searching the web for up-to-date data) because AFAIK it's already more or less a staple in LLM stuff, and his only remarks about is is that it is 'solid but misses sometimes'.
Its because of the em dashes (- is a normal dash, — is an em dash). Very few real people use those outside of writing books or longform articles.
There's also some strange wordings like "back-pocket tests."
It's 100% LLM generated.
What is much scarier is that those "quick reply" blurbs on Android/Gmail (and iOS?) will be able to be trained on your entire e-mail and WhatsApp history. That model will have your writing mannerisms and even be a stochastic mimic of your reasoning. So, you won't be able to even realize a model answered you, not a real person. And the initial message the model is responding to might be written by the other person's personal model.
The future of digital interactions might have some sort of cryptographic signing guaranteeing you're talking to a human being, perhaps even with blocked copy-pasting (or well, that part of the text shows up as unverified) and cheat detection.
Going even a layer deeper / more meta: what does it ultimately matter? We humans yearn for connection, but for some reason that connection only feels genuine with another human. Whereas, what is the difference between a human typing a message to you, a human inhabiting a robot body, a model typing a message to you, and a model inhabiting a robot body, if they can all give you unique interactions?
Everyone who uses a compose key has it available (via ---) — I do. You mean the em-dash though, not the en-dash, and Davidzheng is using hyphens for approximation, not en-dashes.
Not really, as pointed out by others in the thread. Anecdotal of course, but I use em dashes all the time— even in emails and texts (not just long-form writing).
I often write things I want to post in bullets and then have it formulated better than I could by an LLM. But its just applying a style. The content comes from me.
My wife is dyslexic so she passes most things she writes through ChatGPT. Also not everyone is a native speaker.
TBH I've recently felt like that for ~70% of 'top-level replies' in HN, which has slowly pushed me to other mediums (mastodon and discord).
Could just be that the AI 'boom' brought a less programming-focused crowd into the site and those people lack the vocabulary that is constantly used here, who knows.
I'd go out on a limb and say I think probably LLMs made the general population aware of how the "general voice" feels/looks/reads like.
So rather than a lot of people adopting to write like how a LLM writes, the LLM writes as an average of how people been writing on the internet for a long time. So now when you start to recognize how "LLM prose" reads (which I'd say is "Internet General Prose"), you start to recognize how many people are writing in that style already.
I've been in the internet since the early 2000s, I can assure you it does not write like how 'someone on the internet' would write. And when I say that, I mean that for both sides of the internet: it doesn't sound like how 'old school' internet folks would write, but it also doesn't sound like how teens talk either. Neither of these groups write in 'very plain' English regurgitating useless information.
Recent trends/metas in video formats like tiktok and shorts encourage that kind of 'prose', but I haven't seen it being translated into text format in any platform, unless it's written by LLMs.
My point wasn't that it writes like any specific groups, but a general mix-match made up of everyone voice, but a boring average of it, rather than something specific and/or exciting.
Then of course it depends on what models you're talking about, I haven't tried Grok3 myself (which I think you're talking about, since you say "it"), so I can't say how the text looks/feels like. Some models are more "generic" than others, and have very different default prose-style.
Here’s the conclusion of a much more refined initial review by Andrej Karpathy [1] which, I think overall, comports with the substance of my own hot take:
“As far as a quick vibe check over ~2 hours this morning, Grok 3 + Thinking feels somewhere around the state of the art territory of OpenAI's strongest models (o1-pro, $200/month), and slightly better than DeepSeek-R1 and Gemini 2.0 Flash Thinking. Which is quite incredible considering that the team started from scratch ~1 year ago, this timescale to state of the art territory is unprecedented. Do also keep in mind the caveats - the models are stochastic and may give slightly different answers each time, and it is very early, so we'll have to wait for a lot more evaluations over a period of the next few days/weeks. The early LM arena results look quite encouraging indeed. For now, big congrats to the xAI team, they clearly have huge velocity and momentum and I am excited to add Grok 3 to my "LLM council" and hear what it thinks going forward.”
I liked Grok 3 fiction writing style; catches lots of physics of mundane situations such as ringing echo in a closed bathroom we all know well; the prose feels very lively as the result. Kinda like R1 makes situations sharp with details, Grok 3 makes the other way around - rounded by using details.
Well because you explicitly ask it to demonstrate the physics, it came out way too detailed, but point is that it adds details on its own to scenes, make more realistic, not that dry LLama 3.3 style.
Here is the sentence : (She screamed, which echoed off the tile walls. “This is my life now,” she said to her reflection, which looked back at her with a mix of disgust and pity.) Looks good to me. try it on Lmarena.ai.
reply