🤖 working on making computers better to use | Currently building embedded intelligence at https://nitter.cf/t.co/yp7jQwTRZF | Manipulating tensors at https://nitter.cf/t.co/qGY1WlHgk4
Chicago, IL
Joined September 2026
- Tweets241
- Following409
- Followers78
- Likes6K
Coding harness like Claude or Codex are still designed as if the user is carefully monitoring every command line call or code edit. When I come back after an hour to a session these are mostly useless and I have to interrupt the agent to explain what it has done.
A live summary of what an agent has done and still intends to do is the obvious UX here.
In less than half a year, the amount code changes merged by anthropic increased by more than 10x. That is >100x annualized. Expect this to happen across every firm that creates software with a time lag of a couple of months.
ALT From Anthropic blog post about version control.
The source: anthropic.com/engineering/re…
Anthropic experimented with dreaming – an agent reorganizing memory that agents wrote down during their work – in the Spring. OpenAI seems to be also working on dreaming.
From my own experience, this is necessary, underrated, and probably true for any persistent state used by ephemeral agents. I have to explicitly tell an agent to clean my beads setup to get rid of stale beads otherwise I get a bunch of stale bug reports and tasks that have already been done.
As a big Beads user I find this is very exciting. The reason that I use Beads (and why you should use it) beyond giving me tasks lists, memories, etc is that is orthogonal to any model API or agentic harness. So you can avoid being locked into codex or claude code and their proprietary features.
Beads is a version controlled database (dolt) of nodes (beads) linked together. Typically this is used to implement work graphs, detailing what tasks need to be done with links between them (task X depends on task Y). Beads is now adding memories as first class beads, so project data and policies can no be linked to tasks, along with versioning so that a inflight PR can refer to an old version of a policy while a new one can refer to a updated version. Agents can be seen as stateless, so improving the memory system removes the burden on the user to acts as a manual memory system.
Additionally, beads is adding a wire protocol, so an individual bead can link to another bead running on a different database, reducing the need for duplicating shared policies and memories.
ALT Article titled "EXTENDING BEADS: MEMORIES, VERSIONS AND THE WIRE PROTOCOL"
I finally have a in terminal graph viewer for my beads setup. So I can see what my implementor agents are doing without squinting.
🫣I Finally got around to building a TUI for the system that I use to manage agents and generate tasks for evaluations (any Beads fans out there). Just need to polish the UI and find a name for it. Planning on shipping soon.
Incredibly, all the vagueposting from DeepMind researchers about Gemini 4 actually turned on to be true!
Hopefully this isn’t benchmaxed and they fixed the hallucination problem and over confidence of Gemini 3 Pro had, probably from RLing to hard on, uh... The preferences of the median non sophisticated user (which led me to consider it a regression over 2.5 Pro)
Sonnet 5.5 shows that having a larger model is not necessarily more optimal. Parameter count works against RL scaling. OpenAI learned this lesson: they used the 4o pre-train for the base or o1, o3, etc, rather than the larger GPT 4.5.
The smaller the model, the cheaper each RL rollout is. This means that you can do more RL over longer tasks and longer CoT, making use of inference time compute scaling and making up for the lower RL/data efficiency of smaller models by having a wider range of tasks. If you look at the performance graph from Artificial Analysis you see how Sonnet matches Opus. Sonnet 5.5's equivalent performance is achieved through utilizing longer chain-of-thought.
Goodreads as a website is horrible to use. Non functional UI, deadlinks, and a site design from 2005. I can’t image why Amazon is not utilizing it more, especially since it has such a strong network mostly focused on mass commercial literature.
In zero sum games, especially contracts, there is a arms race revolving around who can use AI more effectively. Easy to see how this increases what cusomters are willing to pay for legal services (the price of which was supposed to collapse due to AI).
The Blue Cross/Blue Shield Association says that AI is already driving *up* healthcare costs, as hospitals use the technology to find instances where they can bill more for the same level of care. bcbs.com/about-us/associatio…
The UK government is insane: they are using Gemini!
the uk government's official ai guidance tells civil servants to use gemini flash instead of pro, keep prompts short, and not say thank you to it, for the environment
i do wonder at which point we forked the road. instead of building and using it to automate the work, we write guidance on how to use less of it
we have the attitude of a high school prefect. caring more about the rules than actually doing the work
The name of my $100/month chatGPT plan went from pro to pro lite in the codex CLI, makes sense given the rumors about a new $500 "pro max" plan. What doesn't make sense is that signups for the regular $200 pro plan are still paused. Where are they going to get the extra compute?
The advantage of using DoltLite over SQLite is that the storage engine being content addressable provides it with a natural synchronization protocol. You can compare two replicas using their hashes, walking through the prolly tree to see which chunks have changed.
