HNHacker News
TopNewBestAskShowJobs

mediaman

6,587 karma · joined April 20, 2009

submissionscomments
mediaman··on Blue Origin reaches orbit on first flight of its titanic New Glenn rocket
So there's a tradeoff between image resolution and dynamic range, for a given sensor size?
mediaman··on Magic/tragic email links: don't make them the only option
It’s not about data, it’s customer segmentation. Frequent customers are more price sensitive, and are willing to use the app to get all the discounts, while occasional customers will not, so they can capture both the more price sensitive part of the market while getting higher margins from occasional buyers.
mediaman··on After 3 Years, I Failed. Here's All My Startup's Code
I didn't say it's easy, and it's probably not the right fit for you. But it's at least a plan with a good end-game, which a lot of $10/mo online only business models simply cannot turn into something viable.

50-1,000 person companies aren't actually that big. You can get to know the decision makers and it's not like navigating 20,000 person corporate mazes (also possible if you have the stomach for it).

It does seriously benefit from industry knowledge though. The reason so many people think it's a ridiculous pipedream is because they have no industry experience outside of tech, and that's a big problem -- a solvable one, but most people will never do what it takes to solve it and instead say that it's not possible.

It takes a lot of willingness to do non-tech things though. You can't just set up an online marketing funnel and watch the cash come in. It's relationship driven. The advantage of going after a vertical is that the relationships start feeding each other -- you get social proof within the niche industry. And you don't compete with VC-backed software companies as much because they much prefer horizontal solutions (which can get bigger and achieve bigger exits) versus vertical solutions.

mediaman··on After 3 Years, I Failed. Here's All My Startup's Code
Yes. But it's not for everyone, especially not for a lot of engineers.

Many of these buyers feel more comfortable meeting you in-person, so going to trade shows, writing articles for niche trade publications, and sometimes doing site visits are all very helpful. So is cold calling.

Sales cycles are often slow. You have to get used to customers taking months to decide, having lots of meetings with different stakeholders, and to get good at telling the difference between those who need a lot of meetings but will likely close versus those who will never do anything. It's a very different game than selling subs online, which I know little about.

The product does need to be decent but I disagree with others that it has to be amazing. Most enterprise software is horrible. They will be delighted if you actually solve their problems (rather than the problems horizontally-oriented software vendors want their customers to have), even if it has some rough edges, as long as you're very customer-focused, responsive, and available.

But it also tends to be durable revenue, because you get tightly integrated into their operations if you do it right.

mediaman··on After 3 Years, I Failed. Here's All My Startup's Code
I don't understand why people even think about selling $10/mo licenses.

Just go sell $2-4k/mo SaaS to a medium sized traditional industry that has ancient, crummy software with a target customer employee size of 50 - 1,000 people.

Customer problem discovery is easier, you get reputation network effects within the vertical, you gain trust because you learn the industry lingo, you can actually solve peoples' problems because you're not balancing needs against unrelated industries, they're willing to pay, and once you get 20 customers you're doing pretty OK.

mediaman··on Norway: Taxing Unrealized Gains Has Caused an Entrepreneurial Exodus
Yes, they're just a petrostate. Over half their economy is oil and mining.
mediaman··on Taxing unrealized gains has caused an entrepreneurial exodus in Norway
Norway doesn't care. It is a country with a reputation for good governance and northern-European economic strength. But economically, it is a country that is largely a gas station: like a democratic Russia with more competent governance.

Over half its economy is based on oil and mining. It has failed to develop meaningful economic diversification, and, because it has wisely banked so much of the proceeds of its oil (over US$300k per capita), there's not a lot of pressure to adapt.

Norway will not be a center of innovation anytime soon, except in oil-related fields. Eventually, as oil gets replaced as a source of energy, they may feel more pressure to change. But for now, they suffer from a more sophisticated version of the resource curse.

Sweden is an interesting counterexample, which has a lower GDP per capita but a much more diversified economy. Sweden abolished a wealth tax they used to have almost 20 years ago.

mediaman··on Amazon to invest another $4B in Anthropic
One source [1] puts it at 60-75% of revenue as third-party API, most of which is AWS.

[1]https://www.tanayj.com/p/openai-and-anthropic-revenue-breakd...

mediaman··on Yi Peng 3 crossed both cables C-Lion 1 and BSC at times matching when they broke
Who are you referring to? Putin and Russian oligarchs? If so, how would you imagine the mechanics of prosecuting them to work?
mediaman··on Show HN: Embed an SQLite database in your PostgreSQL table
Why not use jsonb for this kind of thing? Store the schema in one table, one per client, or perhaps one per table per client, and then store the data for that in another table, segregated by customer and table type, with row data stored in a JSONB field using that table's schema.

I normally don't like using JSONB when I could have a rigorous schema, but this sort of application seems reasonable.

mediaman··on The withering dream of a cheap American electric car
I believe it's worth pointing out that with your Toyota, you're still buying an American car. The vast majority of Toyota vehicles sold in America are made in US plants.

Which is a good thing! It shows that those other quality issues are not related to US labor force, or some intrinsic American inability to make high quality goods.

mediaman··on The withering dream of a cheap American electric car
The subsidy was to incentivize adoption of technology to fight climate change. As a result, he took actions that reduced CO2 emissions, which everyone benefits from.
mediaman··on Launch HN: Codebuff (YC F24) – CLI tool that writes code for you
Can codebuff handle larger files (6,000 loc) and find the right classes/functions in that code, or if it finds the file with the necessary info does it load the entire file in?
mediaman··on New York Times Tech Guild goes on strike
Hiring consistently high performing employees is not a solved problem.

Making it hard to fire low performers results in low performing teams, and there are no reliable solutions to this.

mediaman··on Notes on OpenAI's new o1 chain-of-thought models
The API returns how many tokens were used in reasoning, so it would be easy to see any average change in reasoning token consumption. And token prices in general have been extremely deflationary over the last 18 months.
mediaman··on Notes on OpenAI's new o1 chain-of-thought models
The simple answer is: I don't care. I'll statistically figure out what the typical total cost per call is from experience, and that's what matters. Who cares if they lie about it, if the model's cost per call fits my budget?

If it starts costing $1 per call, and that's too high, then I just won't use it commercially. Whether it was $1 because they inflated the token count or because it just actually took a lot of tokens to do its reasoning isn't really material to my economic decision.

mediaman··on Notes on OpenAI's new o1 chain-of-thought models
One trick that people are using, when using Cursor and specifically Cursor's compose function, is to dump library docs into a text file in your repo, and then @ that doc file when you're asking it to do something involving that library.

That seems to eliminate a lot of the issues, though it's not a seamless experience, and it adds another step of having to put the library docs in a text file.

Alternatively, cursor can fetch a web page, so if there's a good page of docs you can bring that in by @ the web page.

Eventually, I could imagine LLMs automatically creating library text doc files to include when the LLM is using them to avoid some of these problems.

It could also solve some of the issues of their shaky understanding of newer frameworks like SvelteKit.

mediaman··on MIT's first freshman class since affirmative-action ban is less diverse
There's a couple observations about these numbers.

The first is that the numbers don't add up to 100%, because students can select more than one race. For the class of 2028 they add up to close to 100%, but they add up to over 120% for the prior year. This suggests that there's been significant changes in how many students select more than one race. It's possible that without an incentive to select an under-represented race, that many fewer students did so.

Secondly, the percentage of white students didn't change much. The portion of students identifying as Asian-descent went up quite a bit, while black/hispanic declined; the rest of the difference appears to be due to the change in multi-race selection.

Overall, this does seem to suggest that the systems previously in place did not really affect white students much, but served to suppress Asian student representation in favor of other minority students.

mediaman··on Solving the out-of-context chunk problem for RAG
For my application we do a land-and-expand strategy, where we use a mix of BM25 and semantic search to find a chunk, but before showing it to the LLM we then expand to include everything on that page.

It works pretty well. It might benefit from including some material on the page prior and after, but it mostly solves the "isolated chunk" problem.

mediaman··on Safe Superintelligence Inc.
Do you believe that no subfactor can ever have a sign opposite of the factor of which it is a component?
mediaman··on From grep to SPLADE: a journey through semantic search
They don't mention BM25, which still outperforms much of semantic search. A fun exercise is to watch the benchmarks of the latest semantic embeddings models and see that they still struggle to match good 'ol BM25.

BM25 uses the relative statistical frequency of words to identify relevant material, along with some adjustments. It doesn't use ML at all, but it works very well, especially for technical content.

SPLADE is capable for some areas but is slow, and often times it doesn't present much of a benefit (or is worse) versus BM25 for technical searches, where specific technical words don't have many synonyms that it would be able to pull.

The best search systems today use a mix of semantic search and BM25 or SPLADE, depending on the type of material and the speed required.

mediaman··on Answering Legal Questions with LLMs
The systems described in the article don’t sound like RAG at all.

RAG systems have a much lower propensity to hallucinate, and generate verifiable citations from the source material.

Though I think they’re better as research aides than “write the final product for me.”

mediaman··on Lessons after a Half-billion GPT Tokens
The industry is troubled both by hype marketers who believe LLMs are superhuman intelligence that will replace all jobs, and cynics who believe they are useless word predictors.

Some workloads are well-suited to LLMs. Roughly 60% of applications are for knowledge management and summarization tasks, which is a big problem for large organizations. I have experience deploying these for customers in a niche vertical, and they work quite well. I do not believe they're yet effective for 'agentic' behavior or anything using advanced reasoning. I don't know if they will be in the near future. But as a smart, fast librarian, they're great.

A related area is tier one customer service. We are beginning to see evidence that well-designed applications (emphasis on well-designed -- the LLM is just a component) can significantly bring down customer service costs. Most customer service requests do not require complex reasoning. They just need to find answers to a set of questions that are repeatedly asked, because the majority of service calls are from people who do not read docs. People who read documentation make fewer calls. In most cases around 60-70% of customer service requests are well-suited to automating with a well-designed LLM-enabled agent. The rest should be handled by humans.

If the task does not require advanced reasoning and mostly involves processing existing information, LLMs can be a good fit. This actually represents a lot of work.

But many tech people are skeptical, because they don't actually get much exposure to this type of work. They read the docs before calling service, are good at searching for things, and excel at using computers as tools. And so, to them, it's mystifying why LLMs could still be so valuable.

mediaman··on Show HN: We built PriceLevel to find out what companies pay for SaaS
What are they going to investigate?

The FTC is going to render contractual terms negotiated in good faith between well-represented and sophisticated commercial buyers “just because”?

If a big buyer wants better terms, the seller might give it to them under the condition the pricing info isn’t leaked, spoiling the rest of their market. Otherwise they don’t give it. So the buyer agrees to confidentiality to get the best terms. There’s zero chance the FTC gets in the middle of this kind of negotiation.

mediaman··on Google CodeGemma: Open Code Models Based on Gemma [pdf]
I think most of them allow for that. Works in vscode and vscode-derived (e.g., cursor) editors.
mediaman··on Google CodeGemma: Open Code Models Based on Gemma [pdf]
Seconding Supermaven here, from the guy that made Tabnine.

Supermaven has a 300k token context. It doesn't seem like it has a ton of intelligence -- maybe comparable to copilot, maybe a bit less -- but it's much better at picking up data structures and code patterns from your code, and usually what I want is help autocompleting that sort of thing rather than writing an algorithm for me (which LLMs often get wrong anyway).

You can also pair it with a gpt4 / opus chat in Cursor, so you can get your slower but more intelligent chat along with the simpler but very fast, high context autocomplete.

mediaman··on Binary vector search is better than FP32 vectors
Generally, it seems that people are starting to see more problems when making vectors of fewer than 1,000 dimensions binary.
mediaman··on Devin: AI Software Engineer
LLMs are quite good at text based tasks such as summarization and extracting entities.

These generally don't require advanced logic or thought, though they can require some moderate reasoning ability to summarize two slightly conflicting text extracts.

Lots of corporate work would be enhanced by better summarization, better information dissemination, and better text extraction. Most of it is pretty boring work, but there's a lot of it.

VC hypes seem to want to mostly focus on fantastical problems, though, which sound impressive at dinner parties but don't actually work well.

If you're a VC, do you want to talk about your investment in a company that finds discrepancies in invoices, or one that self-writes consumer iPhone apps?

Only one of those is actually doable today.

mediaman··on Launch HN: Onedoc (YC W24) – A better way to create PDFs
As a point of reference on pricing, convertAPI charges $0.05 per document conversion at their most expensive tier, and with any level of fixed commitment ($80 - $300 per month) it goes down to $0.016-0.006 per document.

Their PDF conversion is pretty good (I use it for PPT/Word -> PDF conversion), though your product is obviously different and has different/better capabilities for programmatic PDF creation. Still, a reference point.

Pricing page: https://www.convertapi.com/prices

mediaman··on Ask HN: What is the, currently, best Programming LLM (copilot) subscription?
The creator of tabnine also just released supermaven. I’ve liked it so far because it’s fast and holds a lot more of my project in context (300k tokens) so it’s better at using data structures or functions I’ve defined elsewhere.
← PreviousPage 9 of 34Next →