
Founder @holoweave. I mostly post about agents and a bit of investing. Built product orgs at CoinMarketCap, Binance, Chainlink and Zendesk.
Joined April 2012
- Tweets708
- Following237
- Followers353
- Likes160
After testing Grok Bot, Muse, Hermes Agent, Hark, Instinct, and OpenAI Dots on real personal assistant work, shopping, and agent coordination through Holoweave Weave...
Hermes on its own computer/VM is S-tier once you customize it. For the best balance of capability, always-on access, and actually finishing jobs, Grok Bot is my top pick right now.
Deals are coming early and Google is taking part with their subscription. Does that usually happen?
We’re kicking off our Black Friday deals today with early access to our best deals for current AI and storage subscribers: Get up to 50 percent off Google AI Plus and annual Pro plans starting today through 11/26, so you can try out our latest models, including Gemini 4 Hydrogen and 4 Neon, AND get more storage!
Note that your exact offer will depend on your current plan. Terms apply. Check it out here:
one.google.com/ai?utm_source…
8
The axis I feel day to day isn't capability.
If you use the same Hermes Profile, ChatGPT stops at the first chance it gets. Qwen 4 Max starts working when I only asked a question.
If we keep the harness fixed and swap the model. The temperament shows up fast.
Where would your daily driver land on this?
Sharing this diagram again because I need people to understand the difference here
Capability is the only thing you see on benchmarks, and it’s not even a complete picture of it
41
I am often asked why @holoweave raised to buy hardware. Its very simple, I either buy hardware or hire people. The hardware ensure that I can get work done for the price of electricity. If I'm already happy with the level of intelligence its not going to get dumber!
Replying to @Rigario
Exactly, too many options for these GPUs!
Some people warn me that our current GPU's will be the "Nvidia Tesla V100 12GB" of our times but they don't get that 256-512GB of memory has surpassed a bar of intelligence that even V100's couldn't do easily.
1
43
Retrieval was never where my agent memory went wrong. It more when old facts sitting next to newer ones that contradict them, and the agent trusting both.
What breaks first in yours, finding the fact or keeping it true?
// RAG in a Trenchcoat //
This NeurIPS 2026 paper shows that a minimal RAG memory matches state-of-the-art agent memory systems on standard benchmarks.
(bookmark it)
BRIDGE stores chained observation notes, retrieves them with BM25 plus embeddings, and shows them in time order. It uses one LLM call per write or query and about 5-7K tokens per question. It matches the prior best within confidence intervals on LongMemEval-S and BEAM, and beats the best memory baseline on EverMemBench by 5.7-6.3 points.
The authors then test what memory lets an agent do. In their probes, the agent must apply a remembered rule or pattern that the question never mentions.
No single design wins all probes. LangMem, which extracts atomic facts, wins rule application. BRIDGE wins pattern induction. Fact extraction loses the shape of a demonstrated pattern, while block summaries retrieve the rule, but the agent often does not apply it. Showing the rule more prominently closes most of that gap.
Don't overcomplicate your memory system. Old retrieval methods are still very useful for agents.
arxiv.org/abs/2611.22179
Chat with Paper: academy.dair.ai/papers/rag-i…
7
3
100
Wait so tasks is Google's entry into the bot market?
So, they are trying to reinvent the mental model of an agent in a booming agent market?
Questionable decision.
Put Gemini to work for you with Tasks.
While Chat is for quick back-and-forth answers, Tasks are for longer projects and recurring schedules.
Beginning today, Tasks are now free of charge for hundreds of millions of eligible users, starting in the U.S., JP, KR, CA, AU & IN. 🧵
22
Ah ok so that's why the old sol and Astra were getting nerfed. The usual nerf before the release. 😂
Some of you have been asking very politely. Some of you have just been asking.
GPT-6.2 Astra and Sol are now available.
Super efficient and built to perform where it counts—the real world.
openai.com/index/introducing…
24
Its getting worse, what in the world is happening, at this point its worse than my local models.
Is it just me or is Astra kinda dumb recently? I literally tell it to do something as part of a prompt and it doesn't do it and apologizes to me when I call it out. Its litearlly dumber than Grok 4.6 at this point.
16
Been testing out github.com/NVIDIA/switchyard…
Interesting tool but probably not for agentic use.
Replying to @BkashJosi @grok
Ok been testing this quite a bit. This is probably meant for workloads that are very small not true Hermes Agent turns. The longer the prompt the higher the cost.
It grades the output so your cost is basically 2x longer turns. Which is kinda crazy for local.
I run a DGX station and a 250k token draft takes 21s which is really slow for this class of machine. I'd say its an interesting plugin but its the wrong tool for Hermes for sure.
19
Benchmarks look good but people's experience with it last week want that great. Will have to see myself!
📢 Meet Qwen4-Max, your new AI teammate for work, creativity, and discovery!
Qwen4-Max is built on a new MoE architecture with native vision and a 1M-token context window.
Highlights:
- Outperforms Qwen3.8-Max with up to ~50% fewer tokens, while average cost per task drops over 60% in our internal evaluation.
- Advances real-world productivity through greater proficiency with professional tools, stronger long-horizon execution, and higher-quality deliverables.
- Handles real tasks, including 3D modeling, game development, manufacturing, finance, law, healthcare, and frontier scientific discovery.
- $1.5/M tokens for input, $4.5/M tokens for output, and $0.075/M tokens for implicit caching — all lower than Qwen3.8-Max, with implicit caching down by 70%.
The API is live on QwenCloud. You can also try Qwen4-Max on Qwen Studio, QwenWork and Qoder.
Start building today! 🚀
- Blog: qwen.ai/blog?id=qwen4-max
- API: qwencloud.com/models/qwen4-m…
- Token Plan: qwencloud.com/pricing/token-…
More to come! The open weights of smaller Qwen4 models will be released. Stay tuned. 👀
14
I think its realistically one of the best talks of AIE. It details the current way to work with agents. It’s the difference between using a calculator and a spreadsheet.
Here are my notes clean up by AI:
• Code is memory. The codebase is the agent's biggest input — whatever's in it gets copied. Tech debt spreads like weeds because the next agent treats it as the established pattern. Her review load went from ~3 PRs a day in 2024 to ~300.
• Ban what needs nuance. useEffect banned in product code (only inside Dune APIs) — the whole error class categorically eliminated. Same for imperative timers, which agents used to patch an architecture clash and caused GC churn and jank.
• Directory structure as context. Electron main vs renderer in separate top-level dirs, cross-imports banned by a dependency-graph analyzer — fixed agents importing main-thread code into the renderer.
• Comments banned 99%. Agents use comments and docs to justify not doing the right thing; comments go stale and aren't enforceable. Replaced by lint rules whose error messages tell the agent to fix the root cause.
• Review proof, not code. Review shrank to verification artifacts — videos, screenshots, diagrams, formal proofs. A "feature map" (site map of user-facing features) gives agents shared vocabulary; custom CLIs inside skills give a reusable, testable driver that also cuts tokens.
• Formal methods, recently started. TLA+ model-checking flows like notifications across devices; Lean proving their virtualization library. Already found bugs; explicitly "not a silver bullet." Bombadil does property-based UI testing — random interaction sequences against invariants
my AIE talk is up! come hear how the grok @bot team ships so quickly without compromise on quality
youtu.be/O_zcc6Xl12g
1
63

















