If you are running MacOS, I would recommend Agent Safehouse.
Well maintained and is built on existing sandbox-exec so you are not locked in and can always build your own rules independent of the CLI tool.
Agreed, the programmer is not going away. However, I expect the role is going to change dramatically and the SDLC is going to have to adapt.
The programmer used to be the non-deterministic function creating the deterministic code. Along with that were multiple levels of testing from unit to acceptance in order to come to some close alignment with what the end-user actually intended as their project goals.
Now the programmer is using the probabilistic AI to generate definitive tests so that it can then non-deterministically create deterministic code to pass those tests. All to meet the indefinite project goals defined by the end-user. Or is there going to be another change in role where the project manager is the one using the AI to write the tests since they have a closer relationship to the customer and the programmer is the one responsible for wrangling the code to validate against those tests.
This exactly. How many small businesses were running some sort of Access database built by someone who knew just enough to meet their business process needs? "Brenda always orders these parts on Tuesday so make sure there is a Tuesday ordering table."
Eventually some SaaS solution came in and they evolved their business process because of the proposed benefits.
I expect we are going to see a resurgence of good enough software created for the real world hodgepodge of business practices. Not a bad thing in the short term as it will creating adequate efficiency, but long term we are going to reap the technical debt.
My prediction is that the next SaaS evolution is going to be platforms for these random solutions. Just as Salesforce captured CMS market, someone is going to capture the AI agent market where users can write and host their LOB code agnostic to the rest of their infrastructure stack and 'slightly' reduce that technical debt.
This is why I think trail running is so valuable. You must be aware of your changing environment and how to adjust your body's movement to accommodate. You are exercising your mind to calculate your current momentum, intended placement of your next step, and it's potential impact to the rest of your body. You are running your own mental physics simulation as you work the trail, nothing like running on regular roads.
It gave me a "living in the future" feeling the day someone sent me a picture of a phone number through imessage. Barely thinking, I pressed on the phone number in the image and I was prompted to call it. It was like technology and primitive intuition teamed up to create that moment.
There is data but nowhere near the amount of written language that is fairly normalized and without the need to account for additional features such as language, dialect, intonation, facial expression, hand gestures. Speech to text is used as the translation layer as it throws many of those other features away and contextualizes it into a set of tokens that are much more efficient to map between languages.
The reason I like line scan images is because it breaks our mental model of images. We are not looking at the image of a train. We are looking at a time series graph of what occupied a very small defined area in space.
The first definition of this type of procedural generation algorithm was called Model Synthesis by Paul Merrell [1] which built upon texture synthesis. You can even read Merrell's later comparison of the two algorithms [2].
The civilian government agencies spent 248B on contract services in 2023 [1]. Not all of that was professional services, but I expect that we will see an increase in that number as more services are contracted out and a decrease in direct government workforce; a government contractor can still work remotely.
I am not qualified to make any assumptions but I do wonder if a massive investment into computing infrastructure serves national security purposes beyond AI. Like building subway stations that also happen to serve as bomb shelters.
Are there computing and cryptography problems that the infrastructure could be (publicly or quietly) reallocated to address if the United States found itself in a conflict? Any cryptographers here have a thought on whether hundreds of thousands of GPUs turned on a single cryptographic key would yield any value?
Agreed. This was raised within our corp the other week and we read through the privacy and security documentation as it relates to Connected Experiences.
Microsoft has outlined specifically what Connected Experiences covers.[1] [2]
You could argue that predictive text is a product of machine learning but there is no clause allowing for training any generalized large language models using this data. The confusion may have arisen, if they read an article about CoPilot. If the user had a Microsoft Copilot 365 license, then the data would be used as grounding for their personal interaction with CoPilot. But still not used to train any foundational LLMs.
However, even this data is still managed in compliance with Microsoft's data security and privacy agreements.
The article covers this and I think the title is a bit too general. It is a byproduct of how CRISPR works as it targets a specific sequence. In this case the sequence is also present in areas that were non-targeted. Essentially, the sequence was not unique so the process impacted other areas in unintended ways.
It is a case of economic security. US National Security Advisor Daleep Singh spoke about this on the Odd Lots podcast a month ago. The goal is an integrated supply chain across all three countries, and an end result of marketing the ice breaker capacity to other allies.
Its like showing someone a color and asking how many letters it has. 4... 3?
blau, blue, azul, blu
The color holds the meaning and the words all map back.
In the model the individual letters hold little meaning. Words are composed of letters but simply because we need some sort of organized structure for communication that helps represents meaning and intent. Just like our color blue/blau/azul/blu.
Not faulting them for asking the question but I agree that the results do not undermine the capability of the technology. In fact it just helps highlight the constraints and need for education.
I agree and do not think any company would make that investment directly. Nvidia selling to Microsoft renting to OpenAI, I'm sure you could make that add up to $100B on paper. In the long run the economics are likely much more complicated and consist of "agreements worth $x".
Your point is interesting because I see this across the home hardware industry. Cheap electronics were subsidized by data collection and I suppose that market was lucrative until it wasn't. It happened behind the scenes, we knew it was going on, but it was abstracted away enough that we did not think about it. Now what is happening is that the hardware providers are getting into the marketing game directly. Personally in the past year I've given up on two pieces of hardware Alexa and Roku because they have morphed into ad machines. The market is open for privacy, convenience, and quality. Unfortunately, Apple seems to be the only major player in that space right now and we are paying their premium.
The realized value of LLMs is going to be output from a data to AI pipeline. From raw data to actions or insights. This is Microsoft's play to control the entire process. They are attempting to abstract away all the "tooling" needed to manage and process data.
It falls down because there is no discretion left to the consumer. Objectively, the computer is already processing all the data you are interacting with. Subjectively it is assumed its ephemeral. Computer forensics proves this wrong.
So I think to refine my opening statement, the realized value of LLMs is going to be the output from a "curated" data to AI pipeline. Which Microsoft is not providing with this solution.
Another interesting takeaway is that because you are taking a single line capture the pixel distance equates to time not to distance like a normal image.
What that means is that if you pick two points on the car, you can measure that distance in pixels and multiply by the frequency to get change in time.
Divide the actual physical distance of those two points by that result and you know how fast the car was traveling.
https://github.com/eugene1g/agent-safehouse/ https://agent-safehouse.dev/
Originally posted on HN https://news.ycombinator.com/item?id=47301085