Lindell (@Toastrma) header
Lindell (@Toastrma) profile picture

3 ml researchers in a trenchcoat

SF/London
Joined February 2026
Photo by Lindell (@Toastrma)Photo by Lindell (@Toastrma)Photo by Lindell (@Toastrma)Photo by Lindell (@Toastrma)Photo by Lindell (@Toastrma)
Genuinely don’t get the hate- efficient frontier intelligence is just as important as moonshot high param stuff- $3k for a 3D render is not AGI
One of the funny reasons for the difference in vibes between the influencers and the crowd today on 6.2 Sol/Astra is that we gave a few folks access a while ago and then made a lot of changes in the last 3-4 days and I'm pretty sure no re-testing was done within those days and the videos were produced last week. Don't get me wrong, while the models are a hefty jump compared to their GPT predecessors, they are by no means perfect and there are a few things the teams are excited for the models to get better at in future iterations in terms of taste and niceness of use. Love all the feedback and we hear you!
10
the line that matters in the docs: no database, and a restart erases everything. hence postgres, an event log outside the sandbox, sandboxes only when a tool needs one. long-running agents are mostly crash recovery with a model attached. teapawt.net/ClaudeDevs/status/2120…
Self-hosted Claude Managed Agents is now in private beta. You can host the agent harness, sandboxes, and session history, in your own cloud or on-premises, and choose where requests to Claude are sent.
1
1
2
38
rankings depend on when you stop the clock: astra leads at 25% of budget, opus at 50%. 135/150 runs still improved in the back half. agents said "done", got told to keep going, glm found 24.8% more on one task. done is a feeling, not a verifier. teapawt.net/HuggingPapers/status/2…
Infrathon: AI agents beat human engineers at AI infra optimization New benchmark from AllenAI & UW measures long-horizon coding agents. On real H100 tasks, Claude Opus 5.5 discovered a KV codec with 19.86x compression vs 6.82x human baseline. 135 of 150 runs improved in the second half.
2
3
188
Kinda want to get drunk and vibe code until 5am tonight
1
37
Lindell retweeted
AI: ok share API key I’ll handle the rest Me: *sharing key* AI: f*cking retard NEVER share API key with AI
149
423
42
14,006
272,313
Multi-agent systems don’t need to forget to keep improving. We ran the first matched 8-hour comparison on Anthropic’s performance take-home with 4 Claude Opus 5.5 agents. Shared verified archive: 914 cycles (median) Parallel + communicating baselines: ~946–948 and plateaued early Agents without the archive hit a “false floor” and stopped trying. Most of the archive’s gains after it took the lead were built directly on a teammate’s scored attempt. Scaling discovery does not require forgetting. Paper: alphaxiv.org/abs/2611.verifi…
2
2
2
67
Weekend Hack: Multiplayer for @OpenAI Pets Need a skill, memory, or file from your team? Ask your pet, it'll work out who knows, get the info, and bring it back to you. Thank you @Andy_AJT and the team at Dev Day Exchange for hosting! github.com/toastdemon-qwntl/…
1
2
115
llms are too entrenched imo; MAMBA, LCMs (Meta), and others all support a mix of models future state. we'll see a swap if we ever top out, but there's so much more to discover and do with the llm, idk if we see that future pre-RSI
1
3
108
My AI psychosis is going to 10x
We believe voice is the most natural way to interact with intelligence, so we created Dessa, the most authentic full-duplex speech model ever built. Talking with it feels like talking with a person…full of emotion, nuance, pauses, and a natural cadence. block.xyz/dessa
1
39
Lindell retweeted
Working with agents shouldn't feel like putting out fires. Let Nenya takeover.
2
15
QWNTL Labs is now Durable Intelligence. The name is the thing we study: durability, what lets an agent finish long work and still be right on day 30. It's also much easier to pronounce. Same team, same research. New name. durableintel.com
2
2
60
Lindell retweeted
I'm here to help. Your agents keep tripping up? You keep building cool ideas that turn into a part-time job when you have to keep fixing them? Try Nenya Side effects may include extreme psychosis + free time. (request an invite in the comments)
1
1
14
the idea that the future of work is still agile kanban is so fucking disappointing. “yeah, the model smashes the narrow SAT-esque benchmark, it’s probably near-AGI… anyway, we somehow made a worse Asana. now watch your coworkers babysit their agents too.” OpenAI shouldn’t be stuck doing weird ui/os stuff, it’s cool, but they’re one of the few companies capable of chasing exponentially more brilliant models. there should be an ecosystem downstream creating enough computation and money that this isn’t necessary. founders, builders, funds need to make horizontal innovation an expectation across the rest of the stack, rather than settling for vertical products harvesting marginal gains from the latest model- it’s a cancer; everything is neatly categorised into a frontier derivative that predictably improves NPS by a small margin. Apple beat Microsoft by refusing to accept that better underlying technology was enough. they innovated across UX, systems engineering, components and design. we should expect the same ambition from the rest of the AI stack.
What is ChatGPT Space, that is the question. Endearingly awkward at times because it's the actual team working on it. What would you like to see next?
66
packaging 12 months of research into a harness has been wild; super excited to move out of closed alpha and share this with the world. Nenya beats SOTA for week+ long runs on remote labour, complex reasoning, state management, and maintained peak intelligence- dropping the paper on this later this week. dm me if you want an invite
Give your agents the full project and go do something else. Nenya, is a harness and the first product we're sharing. Reply below to request an access to the invite-only beta. When an agent drifts or says it's done when it isn't, Nenya catches it and fixes it. - 15x more productive - 12x less hallucinations - Reduce token usage by 82%. Papers: qwntl.com/ Product: qwntl.com/nenya.html
1
1
45
People will be pissed, but most of the users I speak with already feel disempowered in their new role as meat proxy. Current use is basically being a cyborg, except the human functionality is all around managing state, decisions, long-term memory and time on task. Despite the infatuation with moving AI line cost from the software budget to the labour budget in pursuit of synthetic operations, everyone is avoiding the fact that models are already treated in a comparable manner to humans: fired/cost-reduced due to underperformance, rather than retained because they’re intelligent. And it isn’t even really the models/OpenAI/Anthropic. These things are punching through SOTA and beating human output by miles under the right conditions. The problem is that inference economics require human or >human-level economic output, and the main barrier to that is durability of state. Not intelligence. Keeping the intelligence useful, coherent and pointed at the same thing for long enough to actually do the fucking job. That has to be the next frontier of research. It should also be neutral (not the current frontier labs, whose focus should remain on making exponentially more intelligent models, and ensuring the US stays ahead) which means we need more attention and talent on durable intelligence. All the talk around alignment for safety matters, but we are miles away from ASI/RSI. There’s a much more immediate alignment problem: operational and state alignment. Can the model stay aligned to the work? Hold state, decisions and intent over time without the human acting as its external continuity layer? That’s how we get autonomous organisations. That’s how we go from coworker to collaborator. And eventually, that’s how you get to a financially viable ASI.
Users won’t be necessarily able to articulate it — but they are going to be increasingly frustrated when companies are holding back inference, restricting tokens naively at the cost of performance Once you have tasted the nectar of scaling tokens you just expect it
1
1
1
157