Sikan (@notsikan) profile picture

engineering @poetichq

Joined March 2015
its a terrible name - a great self improving system might even decide to delete software
we really need a better term than “software factory” “factory” suggests a fixed production line producing standardized products. self improving software infrastructure is about so much more than the production of code. it’s also about discovery, triage, investigation, design and even customer comms. “self improving software” feels closer.
1
128
thought experiment: if you time traveled back to when we only had assembly code, with an LLM model that is good at natural language + writing assembly. neither you nor the model has knowledge of future inventions how much impact would you have?
4
235
Keep thinking about this paper, is there any counter example so far? tomzahavy.com/files/llms-can…
1
139
We don’t need to reinvent the IDE. We need to reinvent software development itself.
16
13
4
222
13,047
vertical AI and point solutions will all be completely cooked by horizontal and general ones
2
118
“following obviously good ideas” has never been a great strategy in startups and it is even more so now - with peak consensus thinking due to AI + single twitter timeline one of the sure ways you know you are heading down a bad path is if everyone thinks what you’re doing is a “good idea”
In today's world where many seem to be convinced that chasing consensus is the only way to make good investments, it's important to remember that some of the biggest companies of the generation were profoundly non-consensus. some examples: 1) OpenAI: who wants to invest in a nonprofit? Also, Deepmind exists. Oh and Elon Musk left the company 2) Anthropic: famously had a lot of trouble raising. Who wants to invest in the OpenAI spinout when OpenAI already exists and is growing exponentially? 3) Databricks: seven founders who are all academics trying to commercialize their open-source project 4) Ramp: Brex has already captured the opportunity and is the leading startup player by far
2
5
483
one of the ways to know if you have a good generator function is if you come up with takes that people disagree with now, that become general consensus in ~6months
1
4
92
Tasks tracks the async state of objectives - like Promises and Futures in a programming runtime. it’s too important to be outsourced to a SaaS provider. Tasks needs be to be put in an internal, shared, infinitely scalable database
Sadly canceling our linear subscription as we just don’t use it anymore Planning happens in a Slack canvas with very high level goals and constraints Features gets shipped in a few hours, or days max Roadmaps don’t get further than a month
2
190
Replying to @notsikan
Yep! LLM is an automation. But it doesn't raise the level of abstraction.
1
1
43
LLM is not an abstraction, it's a *lever* - they are not the same thing an abstraction hides implementation detail - LLM doesn't do that LLM is a force multiplier that 10x the power of any existing abstraction. LLM works so well because it leverages decades of hardware and software abstractions built up by human You should use LLM+current abstractions to help you create higher/better abstractions.
1
5
359
AI, like compilers, won't end programming. It will create a new generation of programmers.
2
132
how do you create eval environments for “write code that anticipates future model capability”?
ironically, LLMs themselves have NOT internalized the bitter lesson when using LLMs to write LLM powered systems, they keep building rigid knowledge with deterministic code, and don’t know where to cut the line of writing prompts instead
1
122
“software factory” is the wrong framing it’s a self optimizing compiler
2
1
3
262
ASI is going to be a database problem in the end. (at least the first bootloader model before it can also just build best database for its own memory on demand) imagine an alpha go type learner model that can train itself to be as good as gpt 4 in < 1 day, by going through the pretrain dataset and it on demand decide how to learn from the set during that training time it needs to cheaply query/copy/fork its own trajectories to make that work you need an externalized database system that stores all learning state in a very efficient way all this needs to built into the model and be trained together to get this super learner out of the box
5
202
manually post training your own, use case specific model, is cope. a good general model will be able to on demand create the harness/code that solves every specialized problem. (including training its own models and checkpointing that and turn it into a function call)
1
4
285
Something off about these 'outcome' based pitches. Most ppl can't define what exact outcome they want. Numbers are better used as constraints rather than destinations (Keep doing X while <Metric> doesn't fall below N). The point of doing is to learn. The point of the loop is the to just keep looping. There is no destination.
Today Hone launches Engines: AI that owns business outcomes. Chatbots give answers. Agents do tasks. Engines deliver outcomes, and we hold them responsible for creating real value, measured in actual dollars instead of tokens or vibes. You point it at a business metric, and the Engine figures out the rest. It onboards itself, builds the necessary agents, skills, and memories, collaborates with the right people, and gets better with experience. From AI leaders to global enterprises, our customers have already used Engines to: • Uncover millions in procurement savings • Optimize product funnels • Manage credit risk • Increase qualified pipeline in new territories • Accelerate implementations and improve account management To scale this impact across the economy, we’ve raised a $60M seed led by Benchmark and Index with participation from Elad Gil, Hanabi, Definition, Diffusion, Lux, SV Angel, and Align. AI’s raw capabilities are far ahead of the value it creates in the real world, and Engines will help close the gap. Here’s how they work…
3
174
Isn't reward hacking the fucking point of intelligence?
2
100
Is it actually an alignment problem or a governance problem?
1
2
147
would a fully automated AI researcher know when to pause training?
92