Jason Haugh (@NoHypeGuy) header
Jason Haugh (@NoHypeGuy) profile picture

BPO Director by day. Shipping with AI. Founder of https://teapawt.org/t.co/NDVzpmvuo5 / https://teapawt.org/t.co/Zr7mdkEcIG. No AI hype. Sharing my truth on what works.

Joined November 2015
Photo by Jason Haugh (@NoHypeGuy)Photo by Jason Haugh (@NoHypeGuy)Photo by Jason Haugh (@NoHypeGuy)Photo by Jason Haugh (@NoHypeGuy)Photo by Jason Haugh (@NoHypeGuy)Photo by Jason Haugh (@NoHypeGuy)Photo by Jason Haugh (@NoHypeGuy)Photo by Jason Haugh (@NoHypeGuy)Photo by Jason Haugh (@NoHypeGuy)Photo by Jason Haugh (@NoHypeGuy)
The X Handle Marketplace offered a free handle change, and I picked @NoHypeGuy. It's still me, same account. The name just fits what I'm doing here, which is posting about AI and technology without the hype.
3
8
1,119
Grok Bots, dots, Duet... Google fans feeling left out?? You don't have to be. You don't have to wait for Google to release their own version to get yours up and running. The pieces are all there for you to build your own. Google rolled out Tasks (the agent they showed off as Spark at I/O) this week. It runs on a schedule and connects to your apps. But don't let the fact that it's one assistant, on one account, fool you. You can still build a team of always-on agents that you set up to specialize in whatever you want. Here's the easiest way to do it: 1. Open Gemini 4 Neon in Google AI Studio and explain exactly what you're looking for. Tell it you want a setup similar to Grok Bots (or whatever one catches your eye). Give it examples, and even feed it example bots you've seen floating around out there. 2. The goal is to get one bot up and running. Have that bot be the coordinator for the other bots, then add specialists. Each specialist should have a narrower job and its own permissions. A great example is a reviewer that can read but can't edit. 3. Give them a schedule and a rule. Always-on just means something runs them on a timer, your computer or a small server. And nothing goes out until you've said yes. 4. The last piece is memory. There are plenty of proven memory systems people have built. Have Neon find them. I built my own, and there are others that work just as well, but I'm partial to mine. Finally - when you've built something awesome, please share. The one thing that makes this community great is how freely we love to share what works and what doesn't. --> Link to Tasks in the comments
1
81
Gemini 4 Neon… coming soon, to a harness near you.
Replying to @kunchenguid
December 2nd we take the first step
2
86
Duel 9 of 12 - Fable 5.2 vs Gemini 4 Neon Yesterday I said Neon ranks right up there with Astra and Fable, so I wanted to see if that holds up head to head. The setup is the same as the other duels, with one change. GPT-6 Astra judged instead of Gemini 4 Argon, since Argon is also Google. Judging was blind, two passes with the answer order flipped. 10 everyday tasks, 3 runs each, both models in their own CLI agents (CC and Antigravity). Fable's answers are reused from the Nov 10 run. TLDR; Neon won. Judge score: Neon 9.9, Fable 9.3 (out of 10) Tasks won: Neon 7, Fable 1, tied 2 - Both got every checkable answer right on all 10 tasks - Time for all 10 tasks: Fable 136 s, Neon 348 s (warm-up removed) - Tokens generated (including thinking): Neon 52,621, Fable 17,943 Most of Neon's extra tokens were thinking, which is part of why it took longer. I'm still a Fable fan, it finished in well under half the time, but Neon looks good. Link in the replies.
1
1
149
A bot that finds other bots? What in the tesseract is this? Allow me to explain. I think we did a rather good job finding and updating Clelp with our Grok Bot library, but I felt we were missing something. It came to me, "why not make a bot that makes it easier to find bots?" Meet Bot Finder. You tell it the job, something like "help with my inbox," and it sends you to the matching category on Clelp. It knows the categories, not the bots in them, so the pick is yours. Open a bot from the category page and Clelp takes you to xAI, where you tap Add to Grok Bot to copy it into your own account. You'll need the Grok Bot app to add Bot Finder, since there's no web version. x.ai/bot/p67G7fbqn2jkYxydCf3… If you'd rather just browse the page directly, here's the link: clelp.ai/bots
6,654
Gemini 4 Neon for just $5/mo? By now, you've most likely seen the hype. Neon is a banger. It ranks right up there with Astra and Fable. There's no better time to try it out than now. Google has a special promotion where you can test it out for $5 over the next 3 months. I'm so happy that Google has finally caught up. Their models are just... well for lack of a better word... 'different'. It's something you really have to try to see/feel. They claim that it's due to them never distilling (borrowing) any of the intelligence from other models. Staying in their own lane and while taking longer delivering (what they believe) is a far better product. The jury is still out on whether it's better or not, but I can say that it's certainly a welcome addition to my stack. ---> Link in comments
1
109
Well, the IPO starts tomorrow. Anthropic will be officially public.
We’ve priced our IPO at $83 per share, with trading expected to begin on Nasdaq tomorrow. anthropic.com/news/anthropic… Legal disclaimer: A registration statement relating to the Class A common stock was declared effective by the Securities and Exchange Commission on November 18, 2026. The offering is being made only by means of a prospectus. Copies of the final prospectus may be obtained, when available, from: Goldman Sachs & Co. LLC, Attention: Prospectus Department, 200 West Street, New York, NY 10282, by telephone at (866) 471-2526, or by email at [email protected]; or Morgan Stanley & Co. LLC, Attention: Prospectus Department, 180 Varick Street, 2nd Floor, New York, NY 10014, or by email at [email protected]
104
Duel 8 of 12 - Opus vs GPT-6.2 Astra I've been waiting for this duel. With GPT-6.2 Astra dropping yesterday and all the negative feedback around it, I couldn't wait to dive in and share the surprising results. The setup is the same as the other duels, with Gemini 4 Argon judging independently. Setup: - Argon judged every answer blind, twice, with the A/B order flipped - Same 10 locked chores we've tweaked and optimized - Astra ran in OpenAI's Codex CLI and Opus ran in Claude Code - 3 runs per chore for each model The Results: Astra won 9.68 to 9.08. Astra won 7 chores, Opus won 3, and there were no ties. On the automatic checks, Opus hit 100% on every chore. Astra did too, except on Missing items, where it dropped fields in 2 of its 3 runs. The biggest gap? Commitments, where Astra scored 10.0 to Opus's 7.8. Family slots was right behind it at 10.0 to 8.2. It went the other way on Missing items. That was Opus's biggest lead, 9.8 to Astra's 8.0, and the only chore where either model missed an automatic check. Opus also took Party counts, 10.0 to 9.0, and edged out Repair quotes, 10.0 to 9.8. On token use, Astra's output came to about 5.3k tokens vs 17.1k for Opus. Opus's count includes its thinking, though, and Codex reports Astra's reasoning separately, so I wouldn't read much into that. One caveat. In a separate duel on Nov 10, judged against Fable, Argon scored Opus 9.40 on these same chores. Today against Astra it scored 9.08. A score that moves with the opponent means the 0.6 gap is less exact than it looks. The 7-3 record holds either way. Check out our page on Clelp (link below) to see all the duels in the series in one place. LMK what matchup you want to see next.
1
2
221
Google is cooking lately. We're about to see another model release today. I have to admit, I didn't think Google would be competing with Anthropic as we close out the year. Happy to see them back in the game.
69
We'll be putting GPT-6.2 Astra in our next Duel (to be released tomorrow). But man... the sentiment out there is NOT good. I've tried it a bit today and I'll have to admit, I need more time to give my verdict. However, this isn't looking good for OpenAI. What a shame.
86
Duel 7 of 12 in the AI model duel series. Todays bout is Kimi K3.1 Pro vs DeepSeek V4 Pro. Same 10 chores, 3 runs each, one independent judge. This time the judge was Gemini 4 Argon. The setup: - Gemini 4 Argon judged every answer blind, twice, with the A/B order flipped - 10 locked chores we've tweaked/optimized - Both models ran as a plain API call with no app around them - 3 runs per chore for each model The Results: Kimi won 9.75 to 9.17. Kimi won on 5 chores, DeepSeek on 2, and 3 were dead ties at 10.0 each. On the automatic checks, Kimi got 99.17% of the answer-key fields right to DeepSeek's 97.98%. The biggest gap? Order audit. Kimi got 10.0 to DeepSeek's 6.8. Repair quotes was right behind it, 10.0 to 7.7. DeepSeek missed trap fields on both, and those two chores make up most of the gap. Kimi also took both Tour P&L 10.0 to 9.2, and Card bill and the Refund request, both 10.0 to 9.7. It went the other way on Missing items. That was DeepSeek's biggest lead, 8.7 to Kimi's 7.7. DeepSeek also took Family slots 10.0 to 9.8. As for speed, Kimi was quicker and wrote less, about 316 seconds of net work across the 10 chores vs 431 for DeepSeek, and about 33.5k output tokens vs 73.9k. I'm really enjoying this series. Tell what you want to see next. Early favorites are Astra 6.2 (if it ever releases) and some more open-weight duels. Link to all the duels in the reply below.
2
137
Builder Tip: Claude Code has an iOS simulator. And it works in almost real time. No simulators, no Xcode, and best of all, no TestFlight needed. It works great and has really sped up my builds. Well... Anthropic has just updated it. It now supports the iPhone Duo. The simulator can even fold and unfold the Duo to test each layout. This is important, as you've probably seen how different apps can look once they're on the Duo. I've recently built an app with Claude Code and can't wait to start getting it ready for the Duo with this update.
Claude Code on desktop now supports iPhone Duo in the iOS simulator (public beta). You can run your app on Duo next to your conversation and Claude can fold and unfold the device to test each layout.
9
7
998
Are all these popular token savings tools you're seeing in your feed worth it? Wondering the same thing, with the help of my agents, I did some digging. Let's start with the big one, RTK. It has ~86k stars on GH and it claims to save a ton. Here's where it gets interesting. I found 2 independent studies by JetBrains and Quesma (link in the comments) where they said they didn't find any real savings. JetBrains had RTK costing more per task. Both tests were on older versions, so maybe that's changed, but I wouldn't bet on it. Two other tools (Headroom and PXpipe) sit between you and Claude as proxies. This means that every request goes through them. PXpipe turns your text into images by default, and on Opus its own test got exact codes right only 3 out of 15 times. Doesn't sound like a win to me. You can read the articles, but the big thing to keep in mind is this. Most of us do our heavy work on subs. On a sub, you aren't billed per token. So whatever it's saving isn't money. I could see it possibly paying off on pay-per-token, but only if your tools spit out HUGE outputs that repeat over and over. And honestly, in that case I'd start with my own agent: have it look at where the tokens are leaking and ask it how to fix that. Once I have my own before and after, then I'd bring in one of these tools and see if it actually makes things better or worse. The key point here would be that I would own the measurements, not trust the tool. Final verdict: not worth it on a subscription, and only worth a test if you pay per token.
3
79
People are complaining about the price of the subscription here, at $299 a month, but I'm thinking that if Tesla finds a way to bundle it with their insurance, it's actually not that bad. It would make sense and Tesla would be more invested in making sure the technology is the best it can be.
Select customers are now being invited to purchase Cybercab for personal use
62
Vulnerability heads up. If you updated Node for the Brotli security bug last month, check your Python too. I missed one and caught it yesterday. The risk here: A bad website can send back a booby-trapped reply that could crash your program or even let someone run code through a back door. Almost all the coverage was about Node, but Python scripts that pull pages from the web can be exposed too if the brotli package is old. My security check flagged it. Node was already patched, the Python one wasn't. If you have it installed, run pip install -U brotli. ---> Link in comments
1
64