LoadingAGI

updated fable eval scores

We're all in the same boat with Mythos / Fable guys! Source: Trust me bro!

@nrehiew_ You really believe that? You think they'll not give their researchers access to continue making the ultimate Digital God? How hard is it to spin up 10 laptops with someone else logged in Claude or give API Access with no tracking? Not that hard tbh

So Karpathy, Amanda Askell and Chris Olah all don’t have mythos access now?

Elon and Google real quiet rn with all this Fable 5 stuff going on....

@kimmonismus There are many advantages to be 2nd best it seems...

The next big beneficiary is, of course, OpenAI for two reasons. 1) IPO: OpenAI is concerned that Anthropic would preempt its IPO, resulting in a better valuation. And this was recently a likely scenario. This would have created the image of a second-rate competitor. Now, the question is, how will the ban affect Anthropic's valuation and its upcoming IPO? Who wants to invest in a company that has become persona non grata with US authorities (due to supply chain risk) and may not even be allowed to distribute its best models to enterprises, let alone globally? This will certainly put a significant downward pressure on the valuation. 2) OpenAI has the opportunity to learn from this, to proactively engage in discussions with US authorities to avoid such a disaster in advance, to determine how its model needs to be structured, to obtain the authorities' approval beforehand, and thus essentially use the time to develop a model and secure the necessary authorization to distribute it. OpenAI can learn from this situation and presumably has a better relationship with the US government than Anthropic. Therefore, it was a comparatively successful day for OpenAI. Its biggest competitor suffered a major setback.

@DailyDoseOfDS_ Can someone just copy the harness and plug in other models so we can stop relying on Claude?

Claude Code fully dissected! Researchers from UCL reverse-engineered the leaked Claude source. What they found changes how you should think about agent design. Only 1.6% of the codebase is AI decision logic. The other 98.4% is operational infrastructure. Permission gates, tool routing, context compaction, recovery logic, session persistence. The model reasons. The harness does everything else. This is the opposite of what most agent frameworks do today. LangGraph routes model outputs through explicit state machines. Devin bolts heavy planners onto operational scaffolding. Claude Code gives the model maximum decision latitude inside a rich deterministic harness, and invests all its engineering effort in that harness. The core loop is a simple while-true. Call model, run tools, repeat. But the systems around that loop are where the real design lives: A permission system with 7 modes and an ML classifier. Users approve 93% of prompts anyway, so the architecture compensates with automated layers instead of adding more warnings. A 5-layer context compaction pipeline. Each layer runs only when cheaper ones fail. Budget reduction, snip, microcompact, context collapse, auto-compact. Four extension mechanisms ordered by context cost. Hooks (zero), skills (low), plugins (medium), MCP (high). Each answers a different integration problem. Subagents return only summary text to the parent. Their full transcripts live in sidechain files. Agent teams still cost roughly 7x the tokens of a standard session. Resume does not restore session-scoped permissions. Trust is re-established every session. That friction is the point. The bet behind all of this is simple. As frontier models converge on raw coding ability, the quality of the harness becomes the differentiator, not the model. Paper: Dive into Claude Code (arXiv:2604.14228) We've shared an article on Agent Harness and what every big company is building. Read it below.

@DavidKPiano Lmao I posted about it already. Have to larp the corporate plebs

LinkedIn is going to be *pissed* when they find out about Fable next week

@RetentionAdam Unlimited exports too? Or Unlimited reveals of data?

I’M TIRED, TIRED, TIRED OF DATA COMPANIES BENDING STARTUPS OVER BACKWARDS TO SCREW US ON CONTACT DATA. We’ve bought every vendor under the sun, expensive, cheap, seats, no seats, and every time we switch it STILL feels too expensive. The standard price of contact data is $99 per 5,000 credits. MoltSets.com is $27/mo for UNLIMITED CONTACT DATA. Yes, MoltSets will eventually destroy ZoomInfo and Apollo. YOU: I bet your data quality is shit. ME: TEST OUR DATA COVERAGE AND ACCURACY. It will be just as good as what you’re using. We have to be good to serve RB2B. It was built by me, Tate, and the former president of Apollo/VP product and growth at ZoomInfo. Don’t believe me, test it! YOU: Why are you pricing it so cheap? How can you run a profitable business that way? ME: I’M LITERALLY SELLING SAWDUST. We are using the data team and data asset (very profitably) in 3 other business lines. YOU: Why are you yelling at me? ME: BECAUSE MOLTSETS AT $27/MO UNLIMITED IS FUCKING INSANE. YOU: Aren’t you worried about someone stealing your whole database? ME: Not really. This is a distribution play. If someone stole our database (and I’m sure they will), they will violate TOS (which they will not care about), stop paying us $27/mo (who cares), and then try to do what?!?! Go up against my founder brand, our UGC, and our internal validation engine to acquire customers how?!?!? I think 5x5 already commoditized B2B data, so I want to sell high quality data at a throwaway price and watch an explosion of vibe-coded apps like we've never seen. Imagine what you can build when contact data is (nearly) free! Join the MoltSets wait list here: docs.google.com/forms/d/e/1FAI… Over 2,000 people have already signed up. Beta starts soon.

Everyone having Fable psychosis like GPT 4o had

@wesbos Lmfao. Jokes write themselves. All the deep underlying tech for AI is developed by Canadians. That brain drain went to US leaving Canada without actual powerful model development or talent. If you cant get funding, ppl leave. No matter the push by Carney on Soverign Canadian AI

Canadian LLMs would never do this to you

@ItakGol Regulation

USA has Claude USA has ChatGPT USA has Gemini USA has Grok China has Qwen China has DeepSeek China has Kimi China has MiniMax Europe has?

Agreed. Gemini wont have to worry about this cuz its not even any good

@KashPrime Yeah they won't let you access Fable and other future models. They've shown their cards. Local AI and China will come ahead. You can only go so far if your marketing is to scare everyone with your Digital God

So Canadian academics and tech leaders are foundational to all modern AI tech. The underlying principles were all incubated in Toronto and Montreal. Is America going to ban us from accessing the products of our own tech?

@MiniMax_AI I pray our Chinese brothers release OpenFable at 0.001% cost without nerfs

M3 would never 🙂‍↔️ As a matter of fact, the weights are now open, too. huggingface.co/MiniMaxAI/Mini…

@petergyang Sam has already started to push this with guys from Zoom

Yeah I think ID verification will soon be required to access the best models

@danshipper Just plan on detail with Fable, install codex cli plugin in Claude Code. Tell it to delegate all coding tasks to gpt 5.5 xhigh in multiple threads and review diffs. This saves tokens and works really well

had an idea for a big fable project, set it up, and let it cook came back an hour later and it had triggered the safeguards and fell back to 4.8 10 minutes in back to codex 😬

If you're using @claudeai Fable 5 and find it eating up tokens like crazy, just have it plan and review only -> coding with opus or Sonnet agents

@amasad Why you don't you teach others? Could probably integrate in Replit

I’ve been doing “loops” for a while now. I don’t do much traditional prompting. Most of my prompts are barely a sentence expressing an outcome. - my orchestrator prompts parallel agents - my computer use verifier gives it feedback - my security, production, and SEO agents generate prompts for fixes The industry is typically 3-6 months behind what we’re doing at Replit.

@THArrowOfApollo Code

Clay’s new pricing is probably my fault. We were paying $314 a month, but using (based on their new model) $214,087.50 worth of Clay a WEEK. Here’s the story: A year ago Clay's head of product hopped on a call with me. I told him we were hitting their platform 17.3 million times per week. Almost all custom events (i.e. HTTPs) I remember his response being something close to "Holy shit, I think you are the largest user of Clay" I said yeah that doesn't surprise me. But then it also came up that we were only paying $3,769 a year. We talked about HTTPs, custom integrations, how we were basically using Clay as a giant API orchestration layer. I knew his wheels were turning. If you saw my last post, you know we eventually replaced Clay entirely with a $200/mo Claude Code subscription. 272,000 leads per second vs Clay's 27 hours for the same volume. But before we left, we were the perfect case study for why Clay's old pricing was broken. $314/mo for 17.3 million weekly, for what they now call ‘actions’. Run the math. We were paying $0.00001815 per action. Clay announced their new pricing structure. They split everything into Data Credits and ‘Actions.’ Actions are HTTPs, custom integrations, API calls. The exact things we were doing 17.3 million times a week. The new price per action credit works out to about 1.24 cents each. A 681% price increase for us I know you might say, "But Clay is letting people stay on the old pricing if they want," and I hear you but I also don't know how it makes me feel that someone brand new would have to pay $856,350 per month to get the same advantages I had when I was starting out only 3 years ago. I'm not saying that one call caused the entire restructuring. But I am saying their head of product learned that day that someone was running 17 million HTTPs a week for the price of a nice dinner. And now every HTTP costs 1.24 cents. anyways For the last year, we've been trying to figure out how to get off of our dependency on Clay. That was until Cursor / Claude Code / Codex came out My VP of Growth, @James, who doesnt know how to write a single line of code, touched Claude Code for the first time And three weeks later he replaced Clay for us We could process 272k rows per second now for the cost of a Claude Code sub My last post was about that system Then after that post, Clay announces new pricing that specifically monetizes the exact thing we were doing at a massive scale. Coincidence? Maybe. But I may owe everyone using Clay an apology If your Clay bill just went up, you can probably blame me for that one. Sorry! I put together a system blueprint of what I did to replace Clay for myself -- every tool, the tech stack, a Clay vs custom comparison, and a 6-step playbook for building your own. Plus a video walkthrough where I show you the live system and how each tool actually works. Reply CODE below and I'll DM it to you.

@theojaffee He's making an AI Hedge Fund. Going to cycle that money to AI research and recursive self improvement

It really is impressive how tight a ship Ilya runs. We're coming up on two years with basically zero leaks