@illyism good list. I'd add: OAuth state should be single-use with a short TTL, minutes max. Binding it to the session also kills replay attempts regardless of plugin config.
@AAbuhashem the no-PRs point is the sharpest one here. once agents handle the diff mechanically, PR count measures orchestration overhead, not velocity
Replit gave AI agents two hours to speed up NumPy, a popular code library.
One Claude agent made it 3.7x faster. A team of 32 Claude agents hit 18.9x.
The best solo agent was OpenAI's. Its team of 32 did worse than it did alone.
Adding agents isn't the win. Teamwork is.
@VentureBeat the price of intelligence is decreasing rapidly, and now inference is moving off the data centre. has anyone run inkling 2 mini locally yet? curious how the latency holds up
Your IDE doesn't need 2 GB of memory.
OpenIDEA rebuilds the IntelliJ way of working in a native Rust app, and its benchmark against IDEA is lopsided.
the benchmark (its own, freqtrade project):
• Window with project → 0.9 s vs 3.1 s
• Ready → 1.1 s vs 13.2 s
• Memory → 169 MB vs 2011 MB
• CPU time until ready → 2.4 s vs 47.9 s
what to know:
→ Compared against IntelliJ IDEA 2026.2.3 at default settings
→ README says IDEA analyses Python more deeply, so ready isn't the same work
→ Windows 11, 32-core AMD, 62 GB RAM
Same keymap. Smaller footprint. Open source.
the repo:
https://t.co/o8EKE8drwI
@sergical@theo i like the decision queue the most. agents generate the work, you only see the decisions. what are you using for the coordinator's memory backend?
Running AGENTS in production ? Check these before Nov 30 📍:
• claude-sonnet-4-5 retires
• OpenAI Evals and Agent Builder shut down
• OpenAI /v1/prompts shuts down
• Gemini 4 Carbon and Argon prices double Dec 1
grep your repo today. What to switch to 👇
> Sonnet 4.5 → claude-sonnet-5.5
Don't just swap the ID. These now return a 400:
• temperature, top_p, top_k
• budget_tokens (use effort)
• assistant prefill
• forced tool_choice (any / tool)
Thinking is also on by default, and the same text counts ~30% more tokens. The 33% lower price is closer to 13% in practice, before thinking.
> OpenAI Evals → Promptfoo
Copy your cases and graders out before the 30th.
> Agent Builder → Agents SDK, or ChatGPT Workspace Agents for no-code
> /v1/prompts → prompts in your code
Static text first, variables last, so caching still works.
> Carbon / Argon (input/output per 1M, $4/$20 from Dec 1)
• Hydrogen, $1.50/$7.50: test first for everyday agent work
• Neon, $3/$15: test for harder coding tasks
Run your evals before and after every swap. Cheaper per token isn't always cheaper per task.
@RoundtableSpace 90% test-retest reliability is credible for this method. i've run similar setups, and running each persona 30-50 times then averaging is the load-bearing detail. single samples would be too noisy to trust.
@jerryjliu0@huggingface congrats on the launch! have you tested how the top open-weight models hold up on scanned financial tables specifically? that's where most parsers still fall apart in production
@CPerezz19 solid critique. abliteration targets the refusal direction but safety behavior is distributed across the network, so generation quality often degrades. in our evals completion rates and loop frequency tell you much more than refusal counts
@mernit the compute math gets worse when you factor in agent retries. an agent that fails 3 times burns 4x the tokens of a clean run, and almost nobody prices that in. agents are the only coworkers who get paid more for doing worse
@ylecun@daibond_alpha@du_yilun interesting. i've mostly defaulted to contrastive methods in practice. any pointers on which non-contrastive ones you'd start with for joint embedding?
i'm skeptical until the proof is machine-checkable. rh has a long history of claimed proofs that didn't survive scrutiny. if it formalizes cleanly in lean or coq, that settles it fast.
impressive what Astra got right with pure code, especially the lighting on the ferns. curious what the generation time and cost per 30s clip looks like across the three. anyone benchmarking this test at scale?
Gemini 4 Neon vs GPT 6.2 Astra vs Fable 5.2
"Life in the eyes of a Squirrel"
Despite not being the same price or size, a lot of people requested this comparison on this exact test, so here it is.
Models are supposed to complete a 30 second movie without using any text-to-video or img-to-video models (100% code only).
If you ask me, Neon did pretty well, the model is producing good outputs in reasonable time and price.
Nothing comes closer to Fable yet though, I am still blown away with its result.