calufa

LLMs are hungry for tokens. Every day I see more startups claiming token usage at 50-100k per month. You can always brute-force your way in - just see HuggingFace's recent attacks.

If throwing more money at inference while accumulating compounding technical debt is the new norm, then we are not solving the problem, and the solution space is already covered.

Perhaps there are marginal gains at the expense of quadrillion-params LLM models with 10x the cost and energy. We are simply making inefficiency more expensive, camouflaged by VC money and great marketing.

If that is not plateauing, then I guess I will have to reconsider what plateauing means.

As LLM coding agents plateau— at least for the average engineer without tens of thousands of dollars or swarms of agents to run —I’d say that, from here on it’s going to be about ASICs, specialized LoRA/or-equivalent models, or a Ruby on Rails for LLM context engineering and orchestration, which LangChain and others seems well position, including Google as they own the entire stack. LLM free lunch has been over for a while, perhaps since the ReAct loop, and has been official since Ilya mentioned it at NeurIPS.

I feel the most exciting development these days is self-evolving agents. Especially if you have a way to verify their outputs with a formal system, or with a system developed since the 60s by armies of PhDs.

DeepMinds Gnome is a good example, where they use DFT to verify outputs. Approximating NP-problems is always fun for those who dare.

I am also building in this space. Its a mix between HPC, AI, and hard science. Pretty fun compared to waking everyday to LLM news that seem more like marketing stunts.

Better technology cannot compensate for poor product design. Two weeks' worth of tokens sounds like tens of thousands of dollars by the time it is done. Wouldn't it be better to hire someone who knows what they are doing, get it right, and teach the other engineers why and how? Tech dept and product dept creeping in on every LLM loop, compounding.

Author here. High-level:

- Problem: AI UI generators are high-fidelity by default → teams bikeshed aesthetics before structure is right.

- Idea: use ASCII as an intentionally low-fidelity “layout spec” to lock hierarchy/flow first.

Why ASCII: - forces abstraction (no colors/fonts/shadows)

- very fast to iterate (seconds)

- pasteable anywhere (Slack/Notion/GitHub)

- editable by anyone

Workflow:

- describe UI → generate ASCII → iterate on structure/states → feed into v0/Lovable/Bolt/etc → polish visuals last

It also facilitates discussion:

- everyone argues about structure/decisions, not pixels

- feedback is concrete (“move this”, “add a section”), not subjective

More advanced setups could integrate user/customer support feedback to automatically propose changes to a spec or PRD, enabling downstream tasks to later produce PRs.

clarity, purpose and a path for career growth

Just a heads up, but I tried to signup but the button doesn't seem to work.

Have you looked into tools https://www.multion.ai or https://www.browserbase.com.

Do you have the link and the time in the video where he mentions it?

Congrats on the launch! It’s wild to see AI stepping into spreadsheets like this. Pretty soon there won’t be a part of our workflow AI hasn’t touched.

It's exciting to see all these AI funds and credits, does anyone has a list or website of active funds available? We recently got granted 15k on OpenAI, but would love to get more ;)

Hi,

You can also follow us on Twitter: - https://twitter.com/nius_tv - https://twitter.com/BeyondEarthTV

In April we will be launching our mobile app; a next-gen non-linear TV news aggregator for mobile devices, that converts articles on topics that interest our users into short-form video using AI and trusted media content.

Sign up for our mobile app beta testing program at https://nius.tv.

Thank you, Carlos Chinchilla Founder at NIUS.TV

Carlos here, sole dev and product at the NIUS.TV.

We have taken a similar approach on NIUS.TV. We've even taken it further by synthesizing an AI anchor.

Check it out:

- https://twitter.com/nius_tv

- https://twitter.com/beyondearthtv

The way we synthesize our videos is a little different from Baidu in practice, but similar on a high level.

For video synthesis we use an event-driven architecture using K8s jobs, and a custom job scheduler built in-house that manages all the steps of the video creation process.

Additionally, because GPUs are expensive and knowing that we synthesize our stories in batches through the day, we built a k8s-cluster-builder, that spins up and down the k8s cluster based on pending stories. This saves us a tremendous amount of cost per month, as the cluster only runs when there is actual work for it. As a bootstrapped project, costs are really important.

Our current video synthesis pipeline is comprised of 20+ steps, this includes speech-synthesis, image generation, speech+text aligners (for captions), language models, text generation, among others.

From my experience building the product, and despite the fact that AI is expensive (training wise), it is feasible to build AI-powered products using off-the-shelf DL technology without having large resources.

Hello, I am the sole developer, designer and founder of the project.

We are a two-person company that has bootstrapped the project so far.

I hope you like it. Do not hesitate to ask any questions.

Some technical nuggets:

- We spin up and down a K8s cluster based on load

- We use Tacotron and WaveGlow

- We use a "tweaked" version of Pix2Pix

- We use forced-aligners

- We have a speech-to-text DL model

- We have around ~30 discrete parts including microservices and cloud functions in an event-driven architecture

In broad terms, we manually select news from RSS feeds and then an automated system takes the text out from the news or blog post and generates a video.

I am building an end-to-end text-to-video product that converts text/blog news into TV-like news using multiple AI techs including GANs, Transformers, variational autoencoders, motion graphics, kubernetes, among other techs.

It has taken me some months, but I have CI/CD in place, some traction in twitter, and subscribed users on our home page.

I have bootstrapped the product so far. I'm the solo dev, designer, and creator.

Soon I will have a big upgrade to the AI anchor.

Let me know what you think https://twitter.com/nius_tv

Official home page https://nius.tv

Hi there, I'm the founder, sole dev and designer at NIUS.TV.

I have been working for the last 6 months, 26 hours a day in the product and tech. Now that things are somewhat stable and the vid production pipeline is not breaking due to audio-transcription alignments and other AI-related-corner-cases I wanted to share this with the HN community.

There is still tons of work to do, bugs to iron out, and there are a ton of ideas in the TODO list (exciting stuff).

The goal is to offer an ad-free product that doesn't sell your data while informing you about the topics that matter to you the most. A product that is private, where you can select topics you want to follow, with no social media bluff.

ATM, we are experimenting with Twitter, and we already have a WebApp working at NIUS.TV -- check it out, it is pretty cool. We hope to have an iOS and Andriod version by the end of the year -- fingers crossed.

I hope you like it. If you do, you can follow us or retweet a story. If you don't, please let me know why here https://www.surveymonkey.com/r/N5NHSWW.

:)

-Carlos

Currently it doesn't look at the robots.txt file. I had taken note of this, and will be added in future releases. Thanks for the suggestion.

Here is a sample code of a 1 depth scrape:

https://github.com/calufa/tales-templates/blob/master/core/s...

This is the api call to start the scraper on twitter:

http://localhost:8080/start?process=tales.scrapers.LoopScrap... -template tales.templates.FirstDepthTemplate -threads 2 -namespace com_twitter -baseURL twitter.com

It will basically go through all the links it finds, that can be millions of links. You can also tell it to ignore certain links using regular expressions or via tales java apis.

"... and vector('king') - vector('man') + vector('woman') is close to vector('queen') [3, 1]. You can try out a simple demo by running demo-analogy.sh..."

I downloaded the code but I dont understand how to get queen as a result. I tried submitting king - man + woman, but the bin dont understand it.

For those who want to use a java based solution, I invite you to check out my open source block tolerant (IP Blocking) web scraper that runs on top of aws and rackspace, called Tales. Tales is designed to be easy to deploy, configure, and manage. With Tales you can scrape 10s or even 100s of domains concurrently.

https://github.com/calufa/tales-core

Take a look at Tales.

Tales is a block tolerant web scraper that runs on top of aws and rackspace. Tales is design to be easy to deploy, configure, and manage.

https://github.com/calufa/tales-core

I am just enabling. For me it's not the "what". Its the "who".

this is why jcoulds will be supported soon.

-- well yeah, its a cpu thing, we may need to lower the kb size on the imgs that made up the sprites, and less fps I guess... we need to do more optimization...

maybe some framerate issues, I am aware I need to compress more the imgs to improve it... I will work on that... -- thanks!

...like everything is running slow because of cpu (framerate problems), or is more like the character moves too slow...?

we are 2 friends from Costa Rica trying to build a game engine. The server is made in node.js, all the actionscript code is made by us with no dependency and we sync the sprites and config files using dropbox. -- Feedback is always welcome...

Provisioning is not that hard, nor scaling -- you can always create a image and create servers from that image. In my case, I usually create a server that listens to git and the rest is synced by some program who is also listening the folder where the git repo is.

Heroku looks cool, but I am used to compile things...

I use rackspace and its very cool. Rackspace is fast and simple, and is also cheap. Their server runs fast, and smoothly. AWS is sometimes too complicated, lots of config, lots of choices, lots of weird price changes -- I am talk about the recent spot instances weird pricing fluctuations, and other wierd episodes they have had. Unless you know exactly what you are doing scaling in AWS will take more time and effort -- Unless you are Zynga, and you need a stupid amount of server which almost noone of the server providers except amazon have, or you dont want to deal with relational db scaling problems, or other of the expensive and specific service that amazon offers I dont see why complicate things in the "building stage". AWS is too complex for my use cases, and my software is usually transparent to the server where he is running, so I can always port my code. Rackspace API is stable and it talks JSON, which is cool because my frontend and backend talks javascript.

I guess this question is too open... the answer will always be, depends on what you are building..., and you amount of experience on the field... In my case several 8gb rackspace servers with a cost of 10 dollars a day seems to work well for collecting and crunching gigabytes of data.

Carol Bartz drama was hilarious, but I must agree that she really sucked! its sad because Yahoo devs are making cool things... is just that execs seems focused on building something "better" than AOL, and, as you know just by mentioning AOL things start smelling funny.