serjester

10 ms·

91.
by serjester 2y ago
I think that’s the remarkable thing - even with all of its flaws and its insane pricing, there’s plenty of people that will pay for it (myself included). LLM’s are good at a class of tasks that humans aren’t.
92.
by serjester 2y ago
Synthetic data generation. You can have a really powerful, expensive model create evals so you can tune a faster, cheaper system with similar performance.
93.
by serjester 2y ago
Assuming a highly motivated office worker spends 6 hours per day listening or speaking, at a salary of $160k per year, that works out to a cost of ≈$10k per 1M tokens. OpenAI is now within an order of magnitude of a highly skilled humans wi
94.
by serjester 2y ago
You guys have some of the most complex pricing I’ve ever seen. I tried to read it a couple times and I still have no idea what any of this costs.
95.
by serjester 2y ago
You may want to checkout uv’s workspaces - they’re very handy for large mono repos.
96.
by serjester 2y ago
This is one of the few agent abstractions I've seen that actually seems intuitive. Props to the OpenAI team, seems like it'll kill a lot of bad startups.
97.
by serjester 2y ago
I created an h3 polars implementation and I think one of the big one's is flow modeling. As other's have mentioned, each hexagon has 6 equidistant neighbors. As a result, it lends itself to analyzing telematics data (hence Uber&#x
98.
by serjester 2y ago
Great write up, especially agree on pgvector with small (ideally fine tuned) embeddings. There’s so much complexity that comes with keeping your vector db in sync with you main db (especially once you start filtering with metadata). 90% of
99.
by serjester 2y ago
Never come close to breaking even? You can now get a GPT-4 class model for 1-2% of what it cost when they originally released it. They’re going to drive this even further down with the amount of CAPEX pouring into AI / data centers. I
100.
by serjester 2y ago
This is cool! With that said for anyone looking to use this in RAG, the downside to specialized models instead of general VLMs is you can't easily tune it to your use specific case. So for example, we use Gemini to add very specific al
101.
by serjester 2y ago
Who knows if this leak is credible, but typical SaaS pricing dictates you should be charging 10 - 25% of the value you bring. That implies this agent is expected to deliver 1 - 2M+ of value, essentially replacing a medium size team of knowl
102.
Tips for using Gemini 2.0 for PDF ingestion(sergey.fyi)
58 points by serjester 2y ago|7 comments
103.
by serjester 2y ago
Looking at the graph it seems pretty clear that spending has out stripped revenues for a really long time. This doesn’t seem like something to be proud of.
104.
by serjester 2y ago
Personally I find it frustrating they called it "agentic" parsing when there's nothing agentic about it. Not surprised the quality is lackluster.
105.
by serjester 2y ago
I suppose this was their final hurrah after two failed attempts at training GPT-5 with the traditional pre-training paradigm. Just confirms reasoning models are the only way forward.
106.
by serjester 2y ago
Insane pricing - $75.00 / 1M input tokens & $150.00 / 1M output tokens. They mention it's a big model but it's hard to imagine inferencing costs being 20X higher than 4o. I'm assuming they're doing everythi
107.
by serjester 2y ago
I love this, it seems like 90% of the YC AI startups are geared towards selling to devs. There's so much value to be had finding where AI is relevant in the broader parts of the economy, especially something like construction.
108.
by serjester 2y ago
Good to see more work being done here, but I don't understand why this is tied to someone's proprietary API. Swapping model providers and adding some basic logging is not remotely painful enough to justify onboarding yet another v
109.
by serjester 2y ago
That's essentially what an embedding model is - a smaller, faster model that's good at finding information quickly. Then you feed that to a larger, more powerful reasoning model to synthesize and you've invented RAG.
110.
by serjester 2y ago
We tried something similar and found much better results with o1 pro than o3 mini. RAG seems to require a level of world knowledge that the mini models don’t have. This comes at the cost of significantly higher latency and cost. But for us,
111.
by serjester 2y ago
We've benchmarked a ton of the open models and voyage dramatically outperforms them. I think MTEB is a bad benchmark.
112.
by serjester 2y ago
For some background, Florida is an incredibly challenging insurance market. In 2020, the state accounted for 85% of insurance litigation despite making up only about 10% of total premiums in the US. That combined with sky rocketing reinsur
113.
by serjester 2y ago
I write a lot of Python and personally I find Claude significantly worse than OpenAI’s reasoning models. I really feel like this varies a ton language to language.
114.
by serjester 2y ago
I think at this point, the only question it remains is how Astral will make money. But if they can package some sort enterprise package index with some security bells and whistles it seems an easy sell into a ton of orgs.
115.
by serjester 2y ago
I respect them for being public about this. With that said, this seems quite obvious - the type of customer that chooses Fly, seems like the last person to be spinning up dedicated GPU servers for extended periods of time. Seems much more
116.
by serjester 2y ago
If someone has 0 understanding of what an embedding model is used for, I seriously question their ability to predict winners and loser in AI. > OpenAI has a "text embeddings" API that is used primarily for tasks where you want
117.
by serjester 2y ago
OpenAI is the 6th most popular website in the world and if someone's going to kill Google, it's them. If they IPO'd today, they'd easily be worth hundreds of billions if not more.
118.
by serjester 2y ago
H3 is awesome here! What I don't think many people realize is that H3 cells and normal geographic data (like zips) are not mutually exclusive. You can take zip outlines, and find all the h3 cells within them and allocate your metric ac
119.
by serjester 2y ago
That assumes that you're able to find a model that can match Gemini's performance - I haven't come across anything that comes close (although hopefully that changes).
120.
by serjester 2y ago
The LLM's are near perfect (maybe parsing I instead of 1) - if you're using the outputs in the context of RAG, your errors are likely much much higher in the other parts of your system. Spending a ton of time and money chasing

More ›