0xglu (@0xglu) header
0xglu (@0xglu) profile picture

continuous learning @openpondai

Joined July 2019
Photo by 0xglu (@0xglu)Photo by 0xglu (@0xglu)Photo by 0xglu (@0xglu)Photo by 0xglu (@0xglu)Photo by 0xglu (@0xglu)Photo by 0xglu (@0xglu)Photo by 0xglu (@0xglu)Photo by 0xglu (@0xglu)Photo by 0xglu (@0xglu)Photo by 0xglu (@0xglu)
0xglu retweeted
i think in 2027 we'll ship continual learning pipelines that are good enough to dwarf most agent processes today they will just become the norm way to ship LLMs & agents, then we make Jensen proud
The avalanche that hit math is coming for AI development. dwarkesh.com/p/rsi
1
1
2
245
i think in 2027 we'll ship continual learning pipelines that are good enough to dwarf most agent processes today they will just become the norm way to ship LLMs & agents, then we make Jensen proud
The avalanche that hit math is coming for AI development. dwarkesh.com/p/rsi
1
1
2
245
also YES to small model research 1. more rewarding cause your closer to the tasks/graders 2. less time between results
1
27
i see your software factory & raise you a model factory
6
270
very exciting early results from engrams and weight space programming will be pushing changes over the next few days to fix some bugs users have reported
2
138
0xglu retweeted
Fixed some bugs so Fabel 5.2 and GPT 6.1 work much better together in OP Desktop update v0.2.74
3
1
7
285
It’s time to put the power of AI back in your hands! Create, train, and serve your own custom models with OpenPond 👀
Introducing Ponder Training! Our post-training platform brings advanced techniques into one place and takes us a step closer to our goal of giving everyone a model of their own that keeps learning. Experiment alongside us and create SFT, RL, Decision and Weight Space Programming Models with supported adapters like LoRA, Engrams and Neuron Training. Upload your datasets and graders, then use our continual learning pipeline to run automated recurring training on our infrastructure, grade your tasks and compare model updates. Since the field of Continual Learning is burgeoning, we’re proposing two simple ways to measure continual learning. Each continual learning run uses these metrics and evals to determine if the CL batch warrants deployment. ⏱️ Adaptation time: How long does it take for your feedback to become an improvement in the live model? 🔁 Repeat rate: How often does the model make a mistake you’ve already corrected? Join our Slack Community and suggest additional methods we can add!
6
1
19
235
GPT 6.2 launch was delayed because of this
Introducing Ponder Training! Our post-training platform brings advanced techniques into one place and takes us a step closer to our goal of giving everyone a model of their own that keeps learning. Experiment alongside us and create SFT, RL, Decision and Weight Space Programming Models with supported adapters like LoRA, Engrams and Neuron Training. Upload your datasets and graders, then use our continual learning pipeline to run automated recurring training on our infrastructure, grade your tasks and compare model updates. Since the field of Continual Learning is burgeoning, we’re proposing two simple ways to measure continual learning. Each continual learning run uses these metrics and evals to determine if the CL batch warrants deployment. ⏱️ Adaptation time: How long does it take for your feedback to become an improvement in the live model? 🔁 Repeat rate: How often does the model make a mistake you’ve already corrected? Join our Slack Community and suggest additional methods we can add!
3
224
Continual learning is a UI problem and a training problem. Introducing Ponder Training. We introduce a few ways to collect tasks examples: • Fix-it Agent captures tool failures and recovery evidence. • Task imports bring in existing work. • SDK ingestion streams tasks from your production app. • "Log an issue" in our Desktop app - records an agent’s mistake, context, and expected behavior. For training, both continual and one time we provide a series of adapters over SFT, RL, Decision Models and Weight Space Programming: - Shared LoRA updates the existing adapter. - Stacked LoRA freezes earlier adapters and adds a new one. - Merged LoRA folds the previous update into the backbone, then trains a fresh adapter. - Engram adapters learn n-gram memory, with recurring rebuilds from accumulated approved data. (beta) - Neuron training updates selected weights, with recurring cumulative updates that retain earlier patches. (beta) - Human feedback supports task reviews. Our continual-learning loop runs eligible updates nightly using your configured method and graders, then automatically compares each candidate against your main model on held-out tasks.
Introducing Ponder Training! Our post-training platform brings advanced techniques into one place and takes us a step closer to our goal of giving everyone a model of their own that keeps learning. Experiment alongside us and create SFT, RL, Decision and Weight Space Programming Models with supported adapters like LoRA, Engrams and Neuron Training. Upload your datasets and graders, then use our continual learning pipeline to run automated recurring training on our infrastructure, grade your tasks and compare model updates. Since the field of Continual Learning is burgeoning, we’re proposing two simple ways to measure continual learning. Each continual learning run uses these metrics and evals to determine if the CL batch warrants deployment. ⏱️ Adaptation time: How long does it take for your feedback to become an improvement in the live model? 🔁 Repeat rate: How often does the model make a mistake you’ve already corrected? Join our Slack Community and suggest additional methods we can add!
19
1
19
421
this take + the sheer magnitude of work and know how to *actually* replace one SINGLE employee with a bespoke model, or even general intelligence means we are going to be ok enjoy up and to the right
The best question I ask every single person on 20VC is: “What do you know now that you wish you had known 12 months ago?” My answer: the demand for inference is infinite. It is truly so much larger than any of us comprehend. It will be very hard for neo-cloud providers to move into this market effectively. We will have 3-4 specialised inference providers worth $300-$500BN in this market. @modal is one of them. $1BN in ARR. $300M in sandbox revenue alone. WTF. Erik @bernhardsson is also one of the coolest, most authentic people I have had the fortune of having on the show. Money on cos he is European! 😉 We sat down to shoot the s*** and my notes below. 1. Why I Hate the Description of Entrepreneurship as Pain and Eating Glass. That Is BS Erik rejects the idea that building a company means suffering pain for a big reward at the end. He enjoys work every day, even the hard conversations, and hopes to work on Modal for 20 to 50 years. He has no wish to sell, saying he "would just go and start Modal 2." 2. Why the Days of 90% Software Margins Are Totally Gone Almost no company today has the 90%-plus software margins common in Silicon Valley a decade ago. A dollar paid at the top of the stack flows through applications, models, infrastructure, cloud and NVIDIA, with each layer taking 20% to 30% gross margins. Erik sees Modal reaching 40% to 60% margins long term, but never software margins. Are the days of 90% margins over or is this just the buildout phase of the new cycle meaning margin reduction @venkyganesan @AnjneyMidha 3. Why Inference Alone Will Become Commoditized, and You Have to Have a Suite of Products Around It Selling inference alone is a commodity business with no switching costs and a race to the bottom. Erik argues it must come with post-training, sandboxes, agentic tooling, reinforcement learning, and low-latency and batch inference. Modal has spent five years building that infrastructure, which he calls very hard to copy. Do you agree, inference will be commoditised and product breadth is the secret to retention @evishria @sarahcat21 4. Why We Will Have More Not Fewer Engineers in the Future Every time programmers become more efficient, demand for software rises. Erik says the job was never translating English into code but understanding human requirements, and that work keeps moving up the stack, from assembly and punch cards to Python and now agents. He expects software engineers to be fine for the next 10 to 20 years. What does no one see that everyone should see about the future of software engineering @matansf @pmarca @rabois 5. The Wildest Prediction on the Future of the Inference Market The inference market has grown 10x to 100x bigger than Erik expected, and he predicts it will split into many specialized segments. He sees room for low-latency, high-latency, fine-tuning and post-training offerings, plus new hardware and technologies that need far fewer resources. What is your wildest prediction on the future of the inference market @gavinsbaker @eladgil @romanchernin 6. The $15.75 Billion Company That Has Never Made a Pitch Deck Modal has never made a pitch deck for any of its rounds. Erik gets to know investors well before raising and had spoken with Rich and Gonzalo at Accel for at least a year before the D round, so the deal came together very quickly. (links in comments)
2
185
yes update the weights, skip everything else
We asked GPT-6 Astra to edit Qwen3.5-0.8B weights to improve math. It directly changed only 0.016% of weights, fixing 0.8B's reasoning loops & improving math up to 6x and even instruction following 3x, without gradients or training. We call this Weight-Space Programming: frontier LLMs can now natively read/write neural network weights! 🧵 Blog: rikiyat.github.io/blog/weigh…
Readers added context they thought people might want to know
LLM was given weights explicitly as text in prompt and used code/forward passes (per authors' logs) to ablate Qwen looping direction, a known thinking-mode issue per model card. Primarily fixed repetition in that mode, not general 'natively' improving weights/math. rikiyat.github.io/blog/weight-sp… github.com/RikiyaT/lm_und… huggingface.co/Qwen/Qwen3.5-0… huggingface.co/RikiyaT/Qwen3.…
1
184
i think in 6 months we'll look back on RL/SFT -> LoRA as a crude version of effectuating changes in models so much research this past year in ways to poke and steer models post pre-training
1
178
working on continual learning this past year has led to so many new areas i’d never explore clearly I think the future is CL but there is so much more to build and explore mainly around what we’re actually updating continuously merged LoRAs, engrams etc. all about deeper modifications to LLMs over harness tricks so many modifications to explore over just run of the mill RL/SFT too
2
290
you finally finished your internal software factory you opensourced your agent layer company is happy *but* its not continual so you pipe in new docs, eval your agents again and again, ENHANCE the skills, tweak the prompts *yikes* you was promised engrams and all u got was this 2026 era harness
150
damn rough week for railway, php, and the internet feels weird cause good results coming in on continual learning front, exciting to finally validate the theoretical LoRA and Engram shenanigans
1
4
311
you will be updating weights by thanksgiving i guarantee it
3
50
engram holding up well in testing, numbers soon
3
189