HNHacker News
TopNewBestAskShowJobs

mercurialsolo

226 karma · joined March 22, 2020

Building cabal
submissionscomments
mercurialsolo··on LLMs and Programming in the first days of 2024
Honestly speaking code generation is a form of augmented retrieval. And going further back, I would say human memory is generated from context rather than retrieved (which is why it's often fallible - we hallucinate details).

LLM's today for me are the equivalent of a large scale human memory for code or for faster augmented retrieval - do they hallucinate details, quite often, but do I find it utilitarian versus dragging myself over documentation details - more often than not.

mercurialsolo··on It Sure Looks Like Phones Are Making Students Dumber
One practical example when I was interviewing a candidate a few years back - I don't need to recall or remember anything, I can just google it up at the moment and learn it.

Now this was an extreme behaviour exhibited, but am sure this is the same with everything that we have traditionally been indicators of skill - we still don't understand perfectly the underlying mechanics here of intelligence. Does memory play a role? How does long term memory / encoding affect our ability to have intelligence ? Is our ability to learn something new an indicator of intellect?

Given these unanswered questions - I doubt our current education system really measures intelligence as it has continued to develop over the years because of introduction of devices and an explosion of information.

mercurialsolo··on Figma and Adobe abandon proposed merger
What about this playing a role?

Is the role of anti-trust, big tech regulation and regulatory fees playing a bigger role here?

Is the era of AI more in favour of say both Figma / Adobe competing vs collaborating - similar datasets?

https://edition.cnn.com/2023/12/19/tech/google-play-store-se...

mercurialsolo··on How artists are sabotaging AI to take revenge on image generators
This is just a race to the bottom and creating artificial work - folks creating poisoned data sets and then armies of data engineers who look at creating valid data sets for building models that can be grounded.
mercurialsolo··on Every app that adds AI looks like this
Will the real slim shady please stand up?

The current AI buzz is being ripped out by the wannabes - everything's marketing and buzz and being able to distinguish the genuine jumps forward is harder because we have made everything a wrapper around the same models - u have a diffusion and a transformer model - and u make it easy to call them - what do u get - a 100+ diffusion / transformer wrappers with a library of prompts - voila your new AI app / agent.

mercurialsolo··on I built a plane spotter for my son in minutes
A lot of trivial plumbing work is really being solved by LLM's.

Even in the space of work - We need a lot of plumbing knowledge to get access to a domain e.g. learning about the basics of beekeeping or how to build a nodejs based webapp which makes an API call - a lot of knowledge in a domain is just grind and grunt work which is being automated.

We now need to peer beyond and that's where true exploratory work in any domain lies and is useful.

mercurialsolo··on Entropy in Engineering Processes
We have to delineate between environmental entropy and the entropy within the system. As about adding entropy, every action has effects which can simultaneously both decrease entropy in one variable while increasing in another.

Any system's future state is not entirely predictable. We determine actions and their impact on current entropy while taking decisions.

mercurialsolo··on If you're interested in eye-tracking, I'm interested in funding you
Eye tracking is essentially a model of visual attention. Visual attention is part of the overall attention space and big companies and use-cases are built around visual attention. Today we track attention by explicit interactions, if we can model around implicitly observable interactions - then we have a much larger observable data space around the user.,
mercurialsolo··on n8n.io - A powerful workflow automation tool
Someone should look at a very different take on this space based on cause and effect based modeling.

DX based on cause and effect is a.lot more intuitive for users and reasoning agents

mercurialsolo··on Universal Design Guide Playbook
There's also design centric firms like Airbnb where it does have a seat at the business table. But largely design doesn't get a seat at the table is because the practice hasn't embraced experimentation and science to quantify business outcomes on an ongoing basis. E.g. what is the cost of reducing user complexity in a large system if it's human assisted and you can train individuals to a workflow - some of these have long term business impact e.g. opens up ways for a younger competitor with a better designed product.

Also given that design is largely exposed via the interface it becomes that what works can be easily copied over by a competitor - think Threads vs X.

If we can bring in design to think through on business problems through the consumer adoption and behaviour change lens - it sure will have a much bigger impact than it does today.

Also see some other meta aspects at play - e.g. design is very subjective at the moment and everyone has a choice to what "works for them is good design" .

mercurialsolo··on Code Llama, a state-of-the-art large language model for coding
Is there a version of this on replicate yet?
mercurialsolo··on Chandrayaan-3 Soft-landing [video]
Any space mission is fairly complex no matter the agency private or public.

For me the most interesting part to watch personally was the telemetry. What's the latency involved here, do they have agency in terms of manual overrides and intervention in case things go wrong?

What does RTT look like, are there more efficient encoding of data to allow minimal information?

mercurialsolo··on The Carrot Problem
Business advantages only exist with some degree of information asymmetry. People only want to share this over when either they are not playing the game or the market incentivizes open sharing (e.g. patent systems for innovation).

Unfortunately for business outcomes or in case of war advantages - it has the exact opposite incentive - sharing this leads to loss of market share, more competition, chances of leaks and enemy knowing about your tactics and investing in R&D.

mercurialsolo··on The crash of Air France flight 447 (2021)
This is a fabulously well written article in the means of a technical thriller.

"All of this happened near instantaneously, leaving the pilots completely in control of the airplane with little advance warning."

Complex systems have such a low-window for handling or preventing failure that it is impossible to deal with it at manual scale unless we have a super-good control panels and copilots.

Why we need a lot more work on observability and interpretability while operating complex systems.

mercurialsolo··on GPT Unicorn has drawn a unicorn
What do you think of the current wave of agents spawning off - given that reasoning still needs a lot of grounding?

With the probabilistic models at hands, adding reasoning and planning in some sense falls back to traditional software and system design. Do you see any plausible breakthroughs in this area with agent frameworks?

mercurialsolo··on Paul Graham on Conversations with Tyler
Why do we have not so many great individual players anymore but great teams? Is our rate of progress and build slowing down? Can the brilliance of an individual overshadow big tech?

Why did we have the best quarter in growth for public companies and why is it taking longer private companies to go public?

More than ambition the rules of the game and the frontier has now changed. I don't think ambitious people don't exist - they just realize this is better realized by being within a larger ecosystem and having influence rather than building everything around oneself.

Would we say Hinton was ambitious? Did he have sufficient impact?

mercurialsolo··on An SEO scam: $62k later
I am 100% in agreement here. A lot of the shady practices thrown around by agency operators in the name of getting you leads, and showing you busy work. Social, SEO, PR, all are the same.

Everyone promises a bunch of future results and charges on basis of work done not on outcome.

At the end of the day as an operator and a founder you first need to figure out what works for you and use an agency only for scaling what you know works. Bring the agency for operating and not for strategy at all.

Even for scaling you need to have the playbook dry and cut. If you get the agency to sell you snake oil by promising leads by doing work, or getting you from 0-1 with a channel it never ever works out.

A whole lot of agencies go about doing a bunch of copy-paste work and well designed reports all to keep you chasing mirages that never manifest.

mercurialsolo··on The AI Station: Run LLMs and AI Applications FULLY On-Premise
Is this zapier integrated with LLM's or how do you see this as different? I understand the ability to host LLM's internally and build workflows - but what do you mean by AI apps?

Are these nocode workflows which can invoke LLM's and be exposed as APIs?

mercurialsolo··on Add an AI Code Copilot to your product using GPT-4
- Did you use any fine-tuning with this or is this standard GPT4? - Does quality of response improve with any additional examples? - Does this perform well for specific languages?
mercurialsolo··on People lining up in Argentina to have their iris scanned for Worldcoin
While I would love to hear this from an Argentine's POV rather than speculation on why this is happening.. my guess - this is the closest to get a USD exchange for the local populace without having to go through the regulatory hurdles.
mercurialsolo··on Kenyan moderators decry toll of training of AI
Any form of repeated exposure to content or propaganda is bound to have a lasting impact on both our cognition as well as memory. This is how we are wired. The key is to ensure we are able to draw boundaries - a key reason why we are all talk about social feed / content moderation.

The challenge is when you can afford to pay and get someone else to do it - is it your problem and do you care if it affects someone else.

Do corporations bear responsibility for affecting people who they employ and put into severe conditions which they are aware have long lasting impact?

mercurialsolo··on Patterns for building LLM-based systems and products
Good note on a design pattern for a LLM based product. The biggest focal points will be if we see evolution of frameworks that tackle the hard parts here.

Evals, RAG, Guardrails often times require recursive calls to LLM's or other fine-tuned systems which are based on LLM's.

I would like to see LLM's and models condensed and bundled up into more singular task trained models - much more beneficial versus system design on using LLM's for applications.

This seems like we are applying traditional system design patterns for using LLM's in practice in apps.

mercurialsolo··on Stopping at 90%
Done is a hard concept.

I would honestly like to understand if the 10% is in reality 10% or more like 80%. Is the project done until it has adoption? Until it finds utility in someone else's workflow?

And does it take 10% more work or 80% more because it's adjacent to what you really are good at?

mercurialsolo··on Employer asking to layoff employees using an AI Model – would you do it?
This is absolutely bonkers both by the company as well as employees who they have trusted to build this. This has no positive outcomes on either side.
mercurialsolo··on Launch HN: Tiptap (YC S23) – Toolkit for developing collaborative editors
Are these editors mobile responsive? Or is this something you see more as being used for desktop / cloud based editors?
mercurialsolo··on Why This AI Moment May Be the Real Deal
While life is typically associated with biological systems, intelligence doesn't necessarily need proof of life. A hive, swarm or network can be intelligent as long as it shows self-awareness and has agency.
mercurialsolo··on Why This AI Moment May Be the Real Deal
It's not combining an LLM with data/memory recollection and infinite agency that will lead to leaps in cognition. Complex systems still can do that and still not have self-awareness / cognition. We will need to take some more fundamental leaps to synthesize memory, agency and prediction to operate within a single system where self-awareness resides in the entity and not in the building blocks. LLM's represent a building block towards that cognitive unit, but the real deal is still some ways away
mercurialsolo··on Don’t Forget Intuition: The art of doing science
Intuition and observation - are the cornerstones of the scientific process. While i do believe we don't have shortage of either, it's when the two meet do we have real breakthroughs. We posit theories and new ways of thinking by observation from other places. Often times, that's why scientists tend to have varied interests.
mercurialsolo··on The Long History of Nobody Wants to Work Anymore
Beyond the usual rationale of I don't want to work for pennies, I think hard work as a cultural vibe has fallen out of the times. It's also what we have designed our reward systems around.

Today's reward systems are tuned not around hard work, but around being smart. The disintermediation of media makes it feel like we are all in a celebrity lottery and we just need a lucky break. We don't need to compound our skills or put in the grind.

mercurialsolo··on Can Nature Lie?
Deception can be looked at from the lens of probability & friction. Biological and physical processes optimize for low friction especially when it comes to instant reactions.

The outcome of not lying is typically perceived penalty, loss or some form of high-friction endeavour. There is a probability of being caught in a lie, but then the risk reward is what we balance out in our heads at the moment.

Reinforcement on the risks of lying if ingrained within a system will quite obviously lead to the more truthful behaviour being practiced. Think of how Singapore has a high risk reward for bringing drugs into the country.

If you break down a lie into a combination of risk, reward and friction in the alternative, suddenly you put yourselves into situation where you are able to look at it more objectively and decide on whether the situation would warrant a lie from a biological machine.

← PreviousPage 3 of 6Next →