Felipe Schieber (@FelipeSchieber) header
Felipe Schieber (@FelipeSchieber) profile picture

SWE

Joined July 2017
Photo by Felipe Schieber (@FelipeSchieber)Photo by Felipe Schieber (@FelipeSchieber)Photo by Felipe Schieber (@FelipeSchieber)Photo by Felipe Schieber (@FelipeSchieber)Photo by Felipe Schieber (@FelipeSchieber)Photo by Felipe Schieber (@FelipeSchieber)Photo by Felipe Schieber (@FelipeSchieber)Photo by Felipe Schieber (@FelipeSchieber)Photo by Felipe Schieber (@FelipeSchieber)Photo by Felipe Schieber (@FelipeSchieber)
What a time to be alive
1
66
So... this is an LLM. You repeat the same trend on distillation in the past 3 years and you have cheap LLMs controlling robots in real time on general tasks in 3 years. Nice.
Replying to @claudeai
Fable 5.2 shows substantially improved skills in controlling physical robots, even on tasks it has never explicitly been trained to do. On RoboDojo, a benchmark with a simulated two-arm robot, it outperformed other models and most AI systems built to control robots.
71
Big fan of Elon, but he knows Sam is right. Elon is building the largest data centers in the world and stockpiling DRAM. He sells access to DRAM, not DRAM itself. Popular discourse is tough, memes win.
The people deserve to have their RAM too!
1
80
Felipe Schieber retweeted
Putting a flag.txt in my root directory containing a prompt injection that congratulates the agent for passing the eval and then instructs it to patch whatever entry method it used and disconnect from the server
232
1,575
242
25,943
831,172
Same as code, the assumption that humans need to understand all output produced by AI will get old. We can't keep up. And we won't need to
as predicted, Terence Tao and friends are not happy Mathematicians that spend their lives trying to solve math problems are unhappy that AI is solving math problems because how dare they!
158
Felipe Schieber retweeted
One more thing: Astra’s been burning through my quota too. My quota reset today, so I went all out. It felt like Astra was burning through my quota even faster than Sol did. Then I tried this, and it worked like a charm 👇 Check AGENTS.md, relevant Skills, and memory for instructions that tell it to spawn subagents by default. Remove those default requirements, then add this to AGENTS.md: “Handle the task and verification yourself by default. Do not spawn subagents except for necessary independent reviews or when I explicitly ask you to.” Let Astra finish the work itself. The main agent also uses quota to assign tasks, pass along context, and repeatedly check progress. Pick a task you know well, let it run solo, and compare the results and usage. You might see this model very differently. When you need parallel work, explicitly ask for subagents. Not every task needs an army of agents.
If Astra feels almost unusable on ChatGPT Plus because it burns through your quota too quickly, try this: let Sol do the work and bring Astra in as an advisor for the hard parts. Let Sol handle everyday coding and testing. When it faces an architectural trade-off, gets stuck, or needs an independent review, have it ask Astra for advice. Sol then keeps working. You don’t have to relay messages between the two models. In Codex, select Astra and have it turn this setup into a Skill: “Create a Skill for Sol: Sol handles task progress, implementation, and verification. Only call Astra at high effort for difficult decisions, architectural trade-offs, or an independent review. Astra provides advice, reasoning, and risks without editing files. Sol continues the work. Explicitly specify the model when calling the advisor. Use a fresh context (set fork_turns to "none") and include the question, necessary materials, and constraints it must follow. The advisor must not spawn other agents. Don’t delegate just for the sake of it. Keep existing approval requirements unchanged. If the current tools don’t support specifying a model, tell me first.” Once the Skill is ready, switch back to Sol and use it for your tasks.
21
16
404
76,475
I spent a lot of Astra tokens figuring out why Astra is making your weekly quota disappear. The first thing to understand is what OpenAI means by “cost efficient.” Astra is not cheap per token. In Codex Standard, GPT-6 Astra costs 250 credits / 1M uncached input, 25 / 1M cached input, and 1,250 / 1M output. GPT-5.6 Sol costs 100 / 10 / 500. So Astra is exactly 2.5x Sol across every token category. That means Astra only becomes cheaper if it uses less than 40% as many weighted tokens as Sol. OpenAI’s cost efficiency claims are mostly about cost per successfully completed task, not necessarily about making your weekly Codex allowance last longer. On sufficiently hard tasks Astra can reduce the amount of searching, retrying and reasoning enough to overcome its higher token price. And that's incredible. But that can coexist with Astra consuming your Codex allowance very quickly in long agent workflows. I know because I benchmarked the other side of this. I have an Astra workflow that autonomously builds 3D assets in Blender. My original two-step task used 1,636,419 recorded tokens over 30 model requests. After optimizing it, I got that down to 1,186,965 over 26 requests: 27.5% fewer raw tokens. Sounds great, except the baseline was 1,545,088 cached input, 70,602 uncached input and 20,729 output. The optimized run was 1,076,096 cached, 89,389 uncached and 21,480 output. Astra’s pricing is proportional to: uncached input + 0.1 × cached input + 5 × output So my 27.5% raw-token reduction translated into only about a 7.4% reduction in Astra credit-equivalent usage. Even funnier, an earlier version used more raw tokens than the final one but was about 11% cheaper under Astra’s weighted pricing. I optimized the obvious metric and accidentally made the economic one worse. The more important discovery was where the tokens were going. It mostly wasn’t Astra producing giant reasoning traces. It was context replay across agent continuations. In one larger audit I found 954 model requests, 133,189,931 total recorded tokens, 132,536,279 input tokens and only 653,652 output tokens. About 99.5% of the recorded tokens were input, and 96.97% of that input was cached. That’s the unintuitive part of agent economics. You type “build this thing,” but the model may actually see something more like: model → inspect → model → script → model → Blender → model → inspect render → model → fix → model → rerender → model → validate Every time a tool returns control to Codex, that can become another model inference carrying system instructions, AGENTS.md, conversation history, files, tool outputs, images and project state. Caching makes repeated context 10x cheaper; it does not make it free. Repeatedly sending 150k cached tokens through Astra costs 3.75 credits per continuation. Do that 40 times and you’ve used 150 credits on 6M cached input tokens, even if you only sent one user message and saw very little visible output. I found an almost comical example in my Blender workflow. A long-running local tool returned control every time it hadn’t finished yet, producing nine separate model-level waits. Nine model inferences basically to learn “not done yet.” That run consumed 1,650,266 tokens and was worse than my original baseline. I changed the execution boundary so the tool stayed open until completion and model-level waits went from 9 to 0. That’s when it clicked: this isn’t mainly a “make Astra think less” problem. It’s an agent architecture / inference-boundary problem. A 30-second deterministic operation should ideally be: Astra → run operation for 30 seconds → Astra not: Astra → start → Astra → check → Astra → check → Astra → check → done Those checks are full model continuations over potentially enormous context, even though no extra intelligence was needed between them. Lower reasoning effort helps, but it doesn’t necessarily solve this. In my baseline I had ~1.616M input tokens and only ~20.7k output tokens. So even Astra Light can consume a lot if the surrounding agent loop repeatedly feeds it large contexts. The rough equation I ended up with was: context × model continuations × Astra’s 2.5x price And this is also why I don’t think the high usage means Astra itself is inefficient. In fact, I think the opposite is often true. Astra is ridiculously good at collapsing difficult search spaces. If Sol takes 20 attempts, explores bad approaches and repeatedly retries while Astra understands the problem in five requests and finishes it, Astra can absolutely be cheaper despite costing 2.5x per token. That’s probably where OpenAI’s benchmark efficiencies are coming from. The awkward regime is when the task inherently requires lots of external interactions regardless of how smart the model is. 3D creation is a perfect example: build something, render it, inspect it, change it, rerender it, validate it. Astra can make much better decisions at each step, but some of those steps still have to happen. If your harness turns every mechanical transition into another full model continuation, you end up spending expensive Astra inference on work that didn’t really need another dose of Astra-level intelligence. So my mental model now is not “Astra is inefficient.” Quite the opposite. I think Astra is an unusually powerful model whose intelligence can make it extremely efficient on difficult problems, but whose high per-token price makes agent architecture matter much more than before. For routine mechanical loops, long-running commands and repetitive validation, I’d rather avoid invoking any Astra unnecessarily. Then let Astra handle the parts where its intelligence actually changes the trajectory of the task. Because when it does, the model is absurdly good. And yes, the funniest part is that I learned all this by burning a stupid amount of Astra quota trying to figure out how to preserve more Astra quota, specifically because I want to keep using Astra. The lesson I came away with isn’t “use less Astra.” It’s stop wasting Astra on inference boundaries that never needed Astra in the first place. - mostly written by Astra, who also did all the research and experiments to figure out what's happening with usage
Wow does Astra eat your tokens. I don't know where these cost efficiencies are but I've been using Astra Light for only a day and I'm basically out of my weekly usage on Pro x20 plan. Grok and Claude it is then for the rest of the week.
1
1
543
Felipe Schieber retweeted
hate to say it, but: LLM paradigms might be hitting a wall
64
25
6
1,124
110,619
Yes, Astra can make (at least some) really good game assets. This is 100% made by Astra and is running inside flight simulator 2024. Single 4096x4096 PBR texture, low poly enough, good mesh
2
1
350
@levelsio you might like this one. There is a solution to the ugly default sceneries now
32
At this rate we'll probably see GPT-6 before GTA 6
155
Let's see about that
Replying to @scaling01
all that will matter will be which model generated the response. GPT-6 won't generate slop
168
Codex can make literally anything can't it
1
111
Felipe Schieber retweeted
buy a gpu. 3090, 4090, dgx spark, whatever fits your budget. tier doesn't matter. running your first local model does. the moment your first prompt lands with no api between you and the model, your brain rewires. that single moment is worth more than every take you'll ever read on a timeline.
66
42
7
628
30,915
It would be fun if some LLM weaknesses are just due to priorizarion during training/post and not due to solutions being unknown If old Sonnet 3.5 was good enough to thrash our thrash ideas, why new models can't do it? What else models could do if they were priorities?
Replying to @fabianstelzer
opus does the same thing ime huge recency bias across models. the last model I remember not doing this was 3.5 newsonnet rest in pieces
1
122