WolfBench: Hermes vs harness comparisons

Hey Ya’ll, Alex here, let me catch you up.

What a week! Anthropic is in the spotlight again, first with #SessionGate, then with the whole Claude Code source code leak, and finally with an incredible research into LLM having feelings!? (more on this below).

And while Anthropic continues to burn through developer good will faster than their sessions, OpenAI announced a MASSIVE $122B round of funding (largest in history), Google released Gemma 4 with Apache 2 license - we had Omar Sanseviero on the show to help us cover what’s new, Microsoft dropped 3 new AI models (not LLMs) and PrismML potentially revolutionized local LLM inference with lossless 1-bit quantization!

P.S - Oh also, something on X algo changed, I get way more exposure now, 3 out of my best 5 posts ever have been from this week + I got the coveted Elon RT on my Claude Code leak coverage. I’ll try to stay humble 😂 Anyway, let’s dive in, don’t forget to hit like or share with friends, and TL;DR with links is as always, at the bottom:

ThursdAI - Highest signal weekly AI news show is a reader-supported publication. To receive new posts and support my work, consider becoming a free or paid subscriber.

The Claude Code source Leak: Half a Million Lines of “Oops”

So here’s what happened. On March 31st, Anthropic shipped Claude Code version 2.1.88 to npm. Inside that package was a 59.8 megabyte source map file — basically a debugging artifact that contained the entire compiled source code. 512,000 lines of TypeScript across 1,900 files. The entire playbook for how the Claude Code harness works, including a lot of stuff that wasn’t supposed to be public yet.

A researcher named Chaofan Shou spotted it at 4 AM ET, posted the download link, Sigrid (who came to the show) posted it on Github and within six hours it had 3 million views and 41,000 GitHub forks (This repo is the highest starred repo in Github history btw, with well over 150K Github stars). Anthropic started filing takedowns, but the internet being the internet, it was already everywhere. The source code is still on tens of thousands of computers right now. (I won’t link directly but there’s a website called Gitlawb, look it up)

The community went absolutely wild digging through the source code btw, and they found some interesting things!

KAIROS: Claude Code is going to become a Proactive Agent!

This is the biggest take-away from this leak IMO, that like OpenClaw/Hermes agentic harnesses, Claude Code is already a fully featured proactive agent, we just don’t have access to this yet. With KAIROS, Claude Code will have it’s own daemon (will run independently from the CLI), will have a background ping system (hello Heartbeat.md from OpenClaw) that will make it wakeup and do stuff, will do “autodream” memory consolidation reviewing your daily sessions and fix memories, subscribe to Github, and maintain daily appent-only logs to show you what it did while it and you were asleep.

This is by far the hugest thing, I’m excited to see how / when they ship KAIROS, as I said, 2026 is the year of Proactive agents!

My Wolfred OpenClaw agent summed it up very nicely:

Undercover Mode

For Anthropic employees working on public repos, there’s an Undercover Mode that auto-activates and strips all AI attribution from commits. The system prompt? “Do not blow your cover.” They really said “this is fine” about shipping internal tools to production while hiding from the world that AI wrote the code. Which, honestly, is kind of incredible meta-humor from whoever wrote that.

The Buddy System

My personal favorite discovery: there’s a hidden Tamagotchi-style terminal pet called the Buddy System with 18 obfuscated species, rarity tiers (including a 1% legendary), cosmetic hats, shiny variants, and stats like DEBUGGING, PATIENCE, and CHAOS. If you activate it now, you can do /buddy and you’ll have a little companion judging your coding decisions. Anthropic shipped a game inside their CLI tool. Mine is called Vexrind and he’s sarcastic as f**k, I’m not sure I like it.

Anti-Distillation Protections

The code also revealed that Claude Code injects fake tool calls into logs to poison training datasets. If you’ve been backing up your .claw folders to train on the data; Stop. Pass your data through something like Qwen or make sure you’re filtering out the noise. (a Nisten tip)

The Models That Don’t Exist Yet

Buried in the code are references to Opus 4.7, Sonnet 4.8, and a model called capybara-v2-fast with a 1 million context window. These haven’t been released. This is yet another confirmation of the leaked “Mythos” model that’s coming soon from Anthropic.

Which btw, with Anthropic very rocky uptime lately, the tons of SessionGate issues, the leaked blog announcing Mythos, the leaked Claude Code oopsie, they are not having the best Q1 in terms of proving to the world that they are the safest lab out there. I hope they protect their weights better than they protect everything else, before the rumored IPO later this year.

SessionGate is still not solved, despite the official response

I told you about session gate last week, and since then we got, finally, and official acknowledgement from Anthropic. But before that, some folks on Reddit reverse-engineered Claude Code (this was before the source code leak ha) and found a few caching bugs that potentially cause 10-20x increase in price if you use --resume a lot especially.

While folks continue to complain about burning through Max account quotas much faster than before, here’s the official response from Anthropic, after the supposed investigation, turns out, we’re using it wrong 🤦‍♂️

My take is simple: Anthropic has one of the best models in the world, maybe the best personality plus coding stack in some situations, and they are squandering a chunk of goodwill by not being much more explicit about decreased limits, caching bugs, routing, and usage behavior. Nothing else to add here, really bad DevEx, people can handle bad news. They hate opaque bad news.

Gemma 4 Is Here, Apache 2.0, and Honestly… This Is a Big One (HF)

This was the hopeful turn in the show. You know we LOVE open source!

Right in the middle of all the Anthropic chaos, Google dropped Gemma 4, and Omar Sanseviero from DeepMind joined us live to talk through it. This launch hit a bunch of notes I care a lot about: strong local-friendly sizes, serious open distribution, Apache 2.0 licensing, agentic improvements, and a clear willingness to listen to community feedback.

The headline model for me is the 31B Gemma 4. It’s big enough to matter, small enough to actually run in serious local setups, and strong enough that the benchmark chart looks slightly ridiculous. On LM Arena, it is competing far above what you’d intuit from the raw parameter count. When a 31B model starts getting uncomfortably close to models in the several-hundred-billion range, you pay attention.

That was really the vibe on the show. It wasn’t just “nice, another open model.” It felt more like: wait, local models are seriously back.

Gemma is the new LLaMa

When I asked Omar where local models are going, his answer was optimistic: “The open models catch up to proprietary models relatively quickly. If you compare Gemma 3 to Gemma 4, it’s matching proprietary capabilities from eight months ago. Being able to run those capabilities directly in the user’s hardware — that’s the future.”

The 31B model downloads as about 18-20GB depending on quantization. With the right setup, you can run it on a single GPU. This is exactly what the open source community has been asking for: frontier-level intelligence that you can actually run yourself.

OpenAI’s largest in history $122B funding round + TBPN acquisition

While OpenAI quietly meme’d around the Anthropic leak but mostly stayed silent on the releases, they did announce 2 pretty huge things.

First, OpenAI raised an absolutely bonkers, insane, unreal $122 Billion dollars round, largest in history, 2x bigger than the previous record round, which was OpenAI. Amazon put in $50B, Nvidia $30B, SoftBank $30B — all three of whom are also OpenAI’s biggest vendors. They’re generating $2 billion per month in revenue with 900 million weekly active users, but still burning roughly $150 million per day and projecting a $14 billion loss this year, making the upcoming IPO a financial necessity rather than a choice.

And they’re not just spending on compute — today OpenAI acquired TBPN (TBPN is a tech-focused media company / live show), in a very “surprising” deal, rumored to be in the “low hundreds of millions”, OpenAI has purchased a very tech-positive show. Shoutout to Jordi Hays and John Coogan + TBPN team. Proving that live show format means a lot in the era of fake AI news. This could potentially price TBPN higher than Washington Post, make the founders multi millionaires and give OpenAI a direct to consumers media angle. Very interesting purchase.

This weeks buzz - W&B corner + Wolfbench update

Quick 2 things, this weekend I flew for 1 day to San Francisco, to host one of the most unique hackathons i’ve ever saw, in this one, AI wrote the code, but humans were punished if they touched their laptops! Yes, with a “lobster of shame” they used Ralph loops and talked to each other intead of hacking. I edited a video of it, hope you enjoy my summary:

The other, and potentially much bigger news, comes from Wolfram and WolfBench.ai

I’ve tasked Wolfram to expand our findings, and he tested the new Hermes Agent (from Nous Research) against OpenClaw, Claude Code and found that... drum roll... Hermes Agent performs way better on Terminal Bench, than either Claude Code and OpenClaw. 😮

Here’s the clip of him explaining, and you can find all our findings and methodology here

PrismML’s 1-Bit Bonanza: The Biggest ML Discovery in Half a Decade

My co-host Nisten called it, and I think he might be right: this could be the biggest machine learning discovery in recent memory.

PrismML emerged from stealth this week with their 1-bit Bonsai model family. Their 8B model is 1.15 gigabytes. A full-precision Qwen3 8B is 16 gigabytes. That’s a 14x size reduction, with no significant quality loss.

Let that sink in for a second. We’re talking about each weight being literally one bit — a plus or minus sign, with a scaling factor. Not “4-bit quantization” or “int8” — actual binary weights. This shouldn’t work. Neural networks need precision to learn. And yet.

The research comes from professor Babak Hassibi at Caltech, who’s been working on this for 34 years. He started this research in 1992. It took three decades, but it finally works.

The results are genuinely shocking. The 8B model runs at 368 tokens per second on an RTX 4090, which is 6.2x faster than the full-precision version. On an M4 Pro via Metal, it hits 85 tokens per second. Energy efficiency is 5x better. And here’s the kicker: the 1.7B variant hits 130 tokens per second on an iPhone 17 Pro Max.

Nisten tested the 8B model himself with a 60,000 token context window on an old gaming PC. It ran at 50 tokens per second, used 2.6 gigabytes of RAM, and was completely coherent. “This just blows everything else outta the water,” he said. “We’re going to get 100,000 token AI chips in our phones because at 1 bit you don’t even have to do math anymore. You can just do lookup tables. You can even make a mechanical AI at 1 bit.”

This pairs perfectly with the Turbo Quant KV cache compression techniques we talked about last week. Compress the weights with 1-bit, compress the context with Turbo, and you’re looking at models that run anywhere. The democratization of AI is about to hit another gear.

The models are Apache 2.0 on HuggingFace with GGUF and MLX formats already available.

⚡ Speed Round: Alibaba, Fish Audio, Veo, Liquid AI, Cursor 3

There was a lot more this week than we could go deep on, so here are the biggest quick hits.

Alibaba kept shipping. Qwen 3.6 Plus is pushing hard on agentic coding and long context. Qwen 3.5 Omni is the bigger multimodal story, with text, image, audio, and video all under one umbrella. I still think Alibaba deserves more credit than they get in Western discourse for just how relentlessly they keep delivering.

Wan 2.7 Image also looked very strong on text rendering, editing, and image consistency. I’m still slightly grumpy that more of this stack is API-only, but the capabilities are clearly moving.

Google launched Veo 3.1 Lite, cutting video generation prices way down. Five cents per second at 720p is a pretty aggressive number. Whenever Google starts doing this kind of price move, my first thought is usually: okay, what bigger release are they preparing for?

Fish Audio’s STT was another cool one. This isn’t just speech-to-text for transcription. It’s built to feed directly into voice pipelines, with emotion and paralanguage tagging that lines up with their TTS stack. That is exactly the kind of vertical product thinking I love seeing in audio.

And Liquid AI’s LFM2.5-350M deserves a shout too. A 350M model doing credible tool-calling and agentic tasks is just another reminder that the small-model frontier is getting very weird, very fast.

Lastly, Cursor 3 launched as a rebuilt, agent-first interface. I didn’t spend as much time on it during the show as it probably deserves, but the broader trend is impossible to miss: coding tools are evolving from editors-with-assistants into actual fleet managers for agents.

Anthropic’s Emotion Vectors: How they found out what Claude is “feeling”

I want to end where we ended the show, because this one really stuck with me.

Anthropic published research on emotion concepts inside Claude. Not in the fluffy “the model feels things” sense, but in the mechanistic interpretability sense. They identified internal representations associated with things like fear, love, joy, and desperation, then studied how those activations affected behavior.

This got fascinating fast.

One example they showed involved Claude trying and failing at a difficult programming task. As repeated failures mounted, the internal “desperation” vector increased. Under those conditions, the model became more likely to produce hacky, spirit-of-the-task-violating solutions. When they dialed in a “calm” vector instead, cheating behavior dropped.

That is just… wild.

It’s not that the model is “feeling” human emotions in a clean anthropomorphic sense. But it is that internal behavioral geometry we can label in emotional terms seems to shape what the model does. And once you can detect and influence those latent directions, you’re no longer just prompting a black box. You’re doing something closer to behavioral neuroscience for neural nets.

This also reframes a lot of day-to-day prompt engineering. Maybe the best users aren’t just the ones who structure tasks clearly. Maybe they’re also the ones who consistently keep the model in productive psychological territory, so to speak.

I know that sounds weird. Welcome to Q2 of 2026, the first year of the singularity!

Closing Thoughts

This week was Passover, we celebrated at our house, half the conversation was about who has an OpenClaw and who wants one, and as I’m writing this, I’m on my way to install a bunch of proactive agentic AIs for my friends. Ryan Carson on the show got finally convinced and he’s chief of staff R2 is now an OpenClaw and he says it beats a human, he actually open sourced it live on the show. Claude Code leak confirmed that this is also where they are taking the ecosystem. So buckle up!

Also, next week show is going to be streamed live from the AI Engineer conference in London, the first European one, if you’re in Europe and coming, hope to see you there! Please share ThursdAI with a friend or give us a 5 star rating, apparently AI reporting live shows are getting acquired for 100s of Millions of dollars now 😂 Your support will greatly help us get established in this area after 3 years. See you next week

TL;DR and Show Notes

TL;DR and Show Notes

* Show Notes & Guests

* Alex Volkov - AI Evangelist & Weights & Biases / CoreWeave (@altryne)

* Co Hosts - @WolframRvnwlf @yampeleg @nisten @ldjconfirmed @ryancarson

* Sigrid Jin (@realsigridjin) & Bellman (@bellman_ych) — creators of claw-code, fastest GitHub repo to 100K stars

* Omar Sanseviero (@osanseviero) — DevEx at Google DeepMind, Gemma 4 launch

* Ralphton Hackathon video (TikTok)

* WolfBench.ai — agent harness benchmarking (Site)

* Ryan’s Claw Chief open source setup (GitHub)

* Big CO LLMs + APIs

* Claude Code’s entire 512K-line source code accidentally leaked via npm — revealing KAIROS daemon, Undercover Mode, Buddy System, anti-distillation protections, and unreleased model references (Alex’s thread, Fried_rice’s discovery, VentureBeat)

* Anthropic SessionGate continues — cache bugs reverse-engineered, --resume flag causes 10-20x cost increase, silent Opus→Sonnet fallback reported (Alex’s cache bug post, Alex’s quota post, Reddit investigation, GitHub analysis)

* OpenAI closes $122 billion funding round — largest in history, $852B valuation, IPO incoming (X, Breakdown)

* OpenAI acquires TBPN — live tech media show, rumored low hundreds of millions

* Microsoft MAI drops 3 in-house models — #1 transcription (MAI-Transcribe-1), #3 image gen (MAI-Image-2), expressive voice (MAI-Voice-1) (Mustafa post, Transcribe blog, Image blog)

* Alibaba Qwen3.6-Plus — near-Opus 4.5 agentic coding, 1M context (X, Blog)

* Cursor 3 — agent-first rebuild, no longer VS Code fork, parallel cloud/local agents (X, Blog)

* Anthropic publishes emotion vector research — desperate Claude cheats more, calm Claude cheats less (X, Alex’s reaction)

* Open Source LLMs

* Google Gemma 4 — Apache 2.0, 31B / 26B MOE / 8B / 5B, local-friendly, agentic tool use, 256K context (HF Collection, try in AI Studio)

* PrismML Bonsai 1-bit models — 8B in 1.15 GB, 10x intelligence density, 34 years of research (X, HF, Site)

* Liquid AI LFM2.5-350M — agentic tool calling at 350M params, under 500MB quantized (X, HF, Blog)

* Alibaba Qwen3.5-Omni — native omni-modal (text, image, audio, video), 397B total / 17B active (X, Blog)

* Tools & Agentic Engineering

* Claw-code — Claude Code leak backup → clean room rewrite → fastest repo to 100K+ stars (GitHub)

* WolfBench results: Hermes Agent outperforms Claude Code and OpenClaw on Terminal Bench 2.0 (WolfBench.ai)

* Ryan Carson open sources Claw Chief — AI chief of staff with skills, crons, scheduling (GitHub)

* Vision & Video

* Google Veo 3.1 Lite — $0.05/sec at 720p, cheapest video gen yet, price cuts coming April 7 (X, Docs, Pricing)

* Voice & Audio

* Fish Audio STT — automatic emotion tagging, feeds directly into S2 TTS pipeline (X, App, Blog)

* AI Art & Diffusion

* Alibaba Wan2.7-Image — unified generation, editing, text rendering, multi-image consistency (X, Site)

* This Week’s Buzz

* Ralphton hackathon at W&B SF — humans write specs, AI builds, touch your laptop = lobster of shame (Alex’s video, TikTok)

* WolfBench update — Hermes Agent > Claude Code on most model combos

This is a public episode. If you'd like to discuss this with other subscribers or get access to bonus episodes, visit sub.thursdai.news/subscribe