Dreams API (@dreamsapi) header
Dreams API (@dreamsapi) profile picture

Using APIs and webhooks integrating large language models

Denver, CO
Joined March 2023
Photo by Dreams API (@dreamsapi)Photo by Dreams API (@dreamsapi)Photo by Dreams API (@dreamsapi)Photo by Dreams API (@dreamsapi)Photo by Dreams API (@dreamsapi)Photo by Dreams API (@dreamsapi)Photo by Dreams API (@dreamsapi)Photo by Dreams API (@dreamsapi)Photo by Dreams API (@dreamsapi)
The Solve.it.com course from @jeremyphoward & @answerdotai is so good I'm telling friends not to use my 15% discount code. Pay the full $500. They've earned it. Week 1 awesome, and they extended registration till next week. Don't miss this. #programming #education
1
2
1,132
FULL EPISODE 00:06:49 - Terrified of Casey 00:09:37 - Rex on Windows 00:13:09 - What is Rex 00:15:26 - libghostty on E Ink 00:33:43 - Who are Terminals for? 00:39:57 - Making Linux Users Mad 00:43:54 - FAA Grounding Outage 00:47:54 - Antiquated Technology 00:54:58 - How ATC works
7
19
4
347
22,030
Replying to @burkelibbey
.... I've been working on a fork for two years. Some big news coming soon
1
1
51
570
Dreams API retweeted
nixpkgs in a browser with unkernel!
trynix.dev/ now boots on unkernel.dev/ by default: Linux compiled to WebAssembly, emulating only the programs. 3-10x faster boot & execution for ANY #nix #nixpkgs package now in your browser cold jj --version: 3.0 s -> 0.35 s python3: 5.2 s -> 1.1 s boot to a shell: 6.6 s -> 3.6 s Thanks @_pi0_ 🔥🔥🔥🔥
4
6
66
2,129
Dreams API retweeted
trynix.dev/ now boots on unkernel.dev/ by default: Linux compiled to WebAssembly, emulating only the programs. 3-10x faster boot & execution for ANY #nix #nixpkgs package now in your browser cold jj --version: 3.0 s -> 0.35 s python3: 5.2 s -> 1.1 s boot to a shell: 6.6 s -> 3.6 s Thanks @_pi0_ 🔥🔥🔥🔥
6
10
4
127
5,508
A compositional harness (like RLM) can give you better returns on RL compute, compared to standard coding agents. During RL post-training, RLMs learn relatively easily how to reshape their own messy observations into compositions of familiar problems for the neural LLM underneath. Compared to conventional append-only harnesses (~ReAct with compaction), Alex isolates a set of "harness abstractions" that *together* enable compositional generalization: context offloading, programmatic tool calling, and sub-agent calling.
Harnesses used to be a crutch for weak LLMs. In 2026 and onward, with stronger models, harnesses now offer a mechanism for composing an LLM’s existing and new capabilities to generalize well beyond their training environments. In our new paper, LM Harnesses are Compositional Generalizers, we explore how various harness abstractions enable LM systems to generalize over task length, task strategies, recursive task depth, and more through composing existing LLM capabilities. These abstractions provide a means for more data-efficient training, especially over long tasks, by keeping trajectories locally in-distribution (LID) — i.e., shaping all individual LLM calls to be in-distribution. For tasks that take a very long time to roll out, this implies that we can even perform surrogate harness training, meaning we train an individual LLM call instead of doing the very long harness rollout, using purely synthetic tasks, with fake tool calls + sub-agents! A common idea floating around is that LLMs will absorb harnesses and make them obsolete. We disagree, and argue that we haven’t sufficiently explored what harnesses can provide. The full paper, logs, codebase, and data are available: alphaxiv.org/abs/2611.harnes… w/ @SOURADIPCHAKR18 @lateinteraction
26
23
289
12,039
I shared a terminal specification proposal with a group of people today including people working on Claude Code and totally forgot its their IPO day so this is the current status of Slack.
84
35
5
3,813
191,569
I've never in my life been busier than the past several weeks, and I apologize if I've been slow with anyone, but the various remaining threads are all converging - and I think y'all will love them.
7
6
2
208
16,468
Dreams API retweeted
Founders who are raising: I'm working with @ericries on seedlist.com. Tell us about your company, your raise and the investors you've already contacted. If you're ready, we'll help with intros through Eric's network. Brand new, so early adopters welcome.
2
4
140
Dreams API retweeted
We finally released solve.it.com to the public. There are way too many things to cover at once, but let me give you the high level value.
7
25
279
27,834
SolveIt is free try ($3 in free credits) so why not give it a go and see why thousands of users are hooked! 😍 solve.it.com/about#pricing
1
1
13
590
Our goal in making Solveit was not to automate our team out of the process, but to help ourselves grow. It has been so helpful to us internally, that my co-founder @ericries said we have a moral imperative to share it more widely. Here it is: solve.it.com
4
8
1
84
18,040
We built a tool for our own internal team at AnswerAI to let us learn, explore, & develop mastery with the help of AI: SolveIt. Today, we’re releasing it to the public for the 1st time. fast.ai/posts/2026-11-16-sol…
80
219
32
1,804
135,563
Small preview of a WIP of a visual AI System builder made for @DSPyOSS native reusable primitives
1
1
3
160
Another preview of Fleet-RLM using dspy.Flex, dspy Decisions, dspy.RLM , Daytona Sandboxes, FastAPI (UI temporary ) github.com/qredence/fleet-rl…
1
2
51
I'm reading RFD4 to prep for a journal club, and reading that + watching this talk from David Baker is giving me so much excitement about where protein design is going and what we'll be able to do soon! youtube.com/watch?v=1mu3x97t…
2
16
104
6,333
How to Run JupyterLab with uv: exploringml.com/posts/run-ju…
1
1
4
273
So proud and grateful to have this support from @gabepereyra @nikogrupen and team at @harvey - long been a fan of their open sharing of contributions in this space like LAB and extremely early ventures into post-training RLMs! As the capability of weight space programming begins to take shape and evals for it pop up (stay tuned!), I believe it will carry the highest odds at changing our notions of optimization and learning the most, and perhaps resolving the bottlenecks of deep learning that we simply accepted so far, like terrible sample efficiency. A ton of work to be done here - watch for it from @rikiyatakehi and @jayden_teoh_ in the coming weeks.
Very cool work from @rikiyatakehi and @lateinteraction's lab at MIT. They show GPT-6 Astra can read the weights of a Qwen model and update them directly to improve performance. In early results MATH-500 jumped from 7.9% to 48.1% and it even wrote a Breakout policy from scratch. The two research questions I'm most excited about are whether this scales from billion-parameter models to trillion-parameter models and which stages of a modern post-training pipeline it could replace. Today the largest open models are post-trained in several stages. Kimi K3 and DeepSeek-V4 train RL specialists and distill them into one model while GLM-5 and MiniMax-M3 run a chain of RL stages. We're proud to support Omar's lab and to be working with them to scale this up to larger models and test it on LAB.
8
12
137
7,096