Jeff Rampe

@swyx @mattpocockuk @trq212 AA just said Terra has no place on the GPT 5.6 Pareto frontier at any effort level. I can imagine SWE-1.7 is a decent sub agent. It is fast but thinks too much. It’s not a great planning partner.

where i'm currently at for Big Boy projects: - sol ultra to plan - fable 5 to critique - sonnet 5/terra ultra/swe 1.7 to ultracode/slop cannon - devin review to review (using kakuna) ~always use a variant of @mattpocockuk's grill-me or @trq212's interview-me to elicit decisions upfront

@EashanSinha @devindesktop Nice work by the Cognition team! SWE-1.7 is FAST! But boy does it think A LOT. Even with caveman speech, that’s a bunch of reasoning tokens. I do like that it detects frustration and says to itself “Don’t say ‘You’re absolutely right!’”

SWE-1.7 is out and is the most advanced model we’ve trained! Right up there with some of the SOTA models for agentic coding at a much friendlier price Try it in @devindesktop and Devin CLI today!

@Vtrivedy10 @BVeiseh @garrytan @Vtrivedy10 While you’re talking Skills and harnesses, anyone have success using Deep Agents with Skills in Windows? It is designed for POSIX path and fails to find Skills or fails to execute them even when found.

@BVeiseh @garrytan skills are great if written well! skills also just ship as part of the harness (ex: Claude Code), so not sure what the boundary we’re describing is filesystems, bash, compaction, context management of large tool calls all also great, depends what we need per agent/task

not hot take🧊 the thin vs thick harness debate is pretty useless and completely misses the nuance of working backwards from a real goal when we build agents the obvious answer is that it all depends on what you’re building! there’s no end all be all principle this is why we advocate so hard for Open Harnesses! You choose, for your task, and optimize as deeply or as shallowly as you need to hit your Pareto mix of perf/cost/latency ex: pls try to build Cursor’s Async Cloud Agents with a thin harness today and lmk how that goes 🙃 it does several rounds of work that goes over millions of tokens, verifies it progressively, uses specific modes for specific tasks, and orchestrates all of this in a harness at the same time if you’re building a local html slides creation agent, you don’t need the full Claude Code harness! Optimize your token/context spend and focus your agent on the task it’s all engineering systems to shape model behavior (cc @ashpreetbedi has great content on this too) so that they do useful work on our behalf prob the best starting way to do that is start light and expand as needed using evals and dogfooding to more reliably solve your task over time I bet most harnesses will still settle on some combo filesystems, use of bash, compaction, offloading large tool calls that clutter context if you’re doing semi-complicated work thin != good good=good and it’s ok to add stuff if your agent gets better at your task

LiteLLM HAS BEEN COMPROMISED, DO NOT UPDATE. We just discovered that LiteLLM pypi release 1.82.8. It has been compromised, it contains litellm_init.pth with base64 encoded instructions to send all the credentials it can find to remote server + self-replicate. link below

@swyx Is it possible to use like a post code Hook in @windsurf to trigger a model change?

@alexalbert__ Agent Skills in @windsurf?

Agent Skills is now an open standard It's been great to see the traction Skills are already getting in the industry and this makes it easier for everyone to build and contribute to them🚀 agentskills.io/home

@oliver_wang2 @samcharrington @twimlai Great interview! Nano Banana is an excellent product! @oliver_wang2 Users would love more precise and consistent capabilities for photo editing and restoration. We need better feedback options to specify what works well and what does not. Give a bi/tri-nary feedback labeling.

I'm excited to share my recent conversation on Inside Nano Banana 🍌 and the Future of Image Generation Models with @samcharrington for the @twimlai podcast. Check it out! twimlai.com/go/748 via @twimlai

@awnihannun And then distillation?

2023 LLM training vs 2025 LLM training

@P2Lahiru @emollick Rumor is Microsoft is bringing Claude models into Copilot experiences.

Hey Claude: "Please create the PowerPoint shared by the high powered management consultants hired by Hamlet after seeing his fathers ghost" That was the only prompt. Loved that Claude made this from the McKinsey Elsinore office (with the right colors!), also that SWOT analysis!

@johnrushx If you have inference workloads that are not time sensitive, use batch processing from OpenAI at 50% the cost.

I've spent $76,082.42 in August 😓 Can you help me reduce this? Contractors: -$10,137.78 Amazon Web Services: -$4,480.12 MongoDB: -$4,251.50 Google Workspace: -$1,913.61 Google Cloud: -$2,134.74 PayPal: -$4,678.55 Exa: -$1,549.00 Cloudflare: -$350.20 OpenAI: -$15,486.83 Anthtopic: -$16,001.22 BunnyCDN: -$170.00 Notion: -$338.00 Figma: -$420.00 Zoho Corporation: -$374.15 Scrshotone: -$146.00 Ghostinspector: -$190.83 DigitalOcean: -$148.26 Imgix: -$808.12 Pinecone Systems: -$61.47 Mailgun: -$96.51 Grammarly: -$144.00 PandaDoc: -$140.00 Jetbrains: -$12.00 Gamma: -$10.00 GitHub: -$14.00 Zoom Video Communications: -$2,434.74 Scrapingbee: -$848.99 Firecrawl: -$175.00 Cursor: -$40.00 Dataforseo: -$100.00 Apify: -$78.00 Devuap LLC: -$19.99 1Password: -$19.95 Statuscake: -$104.48 Atlassian: -$690.00 Perplexity AI: -$204.00 Webshare: -$59.03 QuickBooks: -$75.00 ElevenLabs: -$22.00 Zapier: -$91.13 Apple: -$76.92 Slack: -$137.18 Hushed: -$4.99 StreamYard: -$106.98 Supabase: -$75.00 Microsoft: -$2,319.58 Webflow: -$24.00 Loom: -$15.00 Clerk: -$37.50 Seo Gets: -$29.00 Intercom: -$248.15 Serper: -$1,250.00 Senty Pty Ltd: -$847.27 Hetzner Online: -$864.71 iPostal1: -$9.99 Twilio: -$315.33 Mailjet: -$17.00 Replicate: -$3.50 Crisp: -$540.00 Lordicon: -$16.00 Lovable: -$20.00 Firstpromo: -$84.15 GoDaddy: -$22.19

@vaibhavk97 Have you seen any updated results for LLMs as calculators testing models like GPT-5, Opus 4.1, and Grok 4?

Misc observation: With multiplication in large numbers it is often the case that models get the first 2-3 digits of the answer correct, which is counter-intuitive since the answers are decoded from left to right. Source for the experiment: github.com/vaibhavk97/Ari…

Was curious about the basic arithmetic abilities of Claude 3 Opus against GPTs. Designed a tiny experiment over the weekend and the results are surprising. Opus is much better than GPTs with numbers! Let's dive into the results 🧵

@dave_alive @emollick Yup, what Dave said. You have to nudge it to get the full GPT-5 model.

@emollick If it's using the same router as chat GPT, telling it to "think very deeply about this" should send you to the thinking one... Or at least that's what it did at first before the latest picker changes. I'm not sure if it still does because the router seems pretty good now

Does Microsoft Copilot use the same GPT-5 router as OpenAI does? I can't get their "GPT-5" to pass me to any good model unless it is explicitly a coding or math task, with no indication of which model I get, which makes the quality of outputs feel very uneven in confusing ways.

gpt-5 is showing improvement over other models in successful completion of longer duration tasks. Granted, the confidence interval is wide: 9 min to 1 hr 4 min. Vending Bench is practical way to evaluate this. Looking forward to the updated benchmarks from @andonlabs

@NickADobos @mckaywrigley Interesting that ‘model-router’ is available in Azure OpenAI API but not directly through OpenAI. learn.microsoft.com/en-us/azure/ai…

@mckaywrigley I think you can override the model router in the app by choosing gpt-5-thinking And it’s the auto-router is not available in the api, you gotta route manually in that case So easily skippable!

My honest GPT-5 review: - It is a *phenomenal* everyday chat model - I will default to it for all normal chats - API pricing is incredible, major points here But code? I will still be using Claude Code + Opus.

@handatalks @mdancho84 Can you run LlamaParse locally? I assumed you had to use their API, but a local option would be great!

@mdancho84 Just use llamaParse

🚨 BREAKING: Microsoft launches a free Python library that converts ANY document to Markdown Introducing Markitdown. Let me explain. 🧵

@emollick Amazing! Can this be the "otter on a plane using wifi" of video gen models?

"[video game] as a community theater production" may be one of the most delightful Veo 3 Fast prompts Please enjoy, in order: GTA, Pokemon, Mario Kart, The Witcher 3, Stardew Valley, Tetris, Mortal Kombat, The Sims, & Death Stranding(!) Yes, the whole prompt was the one above.

@jxmnop @grok combine this idea with the Farseer scaling law paper. arxiv.org/pdf/2506.10972 How big would an optimally trained model be with “The Internet“ as training data?

new blog: How to scale RL to 10^26 FLOPs everyone is trying to figure out the right way to scale reasoning with RL ilya compared the Internet to fossil fuel: it may be the only useful data we have. and it's expendable perhaps we should learn to reason from The Internet (not just math and code)

@Yampeleg @jxmnop @grok explain FSDP zero-3 and explain why Yam is right

@jxmnop FSDP autocorrect correcting to words i frequently use lol

two or three years ago, this was a prevailing sentiment “training the largest LLMs is very hard. only a few people know how. everyone else is failing” it was hard to avoid huge loss spikes. now pretraining is a solved problem. what changed? do we just clean our data better?

@AdamWintle @tg_bytes Hours of great content on their YouTube channel: AI Engineer. @aiDotEngineer

@tg_bytes Great breakdown! I hadn’t heard of this conference before!

Got back from AI Engineer World's Fair in SF and I'm still buzzing. Here's what blew my mind 🧵