• Welcome to AppraisersForum.com, the premier online community for the discussion of real estate appraisal. Register a free account to be able to post and unlock additional forums and features.

A.I. Stuff

I don't know what changed but it has been brutal using any AI service The last two days. I've been shut down multiple times by six different AI over the last 2 days even on paid plans for excessive demand. Anyone else saying this?
 
I think when they come out with a new version, as the case with Claude's Fable 5, they give it highest priority. I am still on my free Fable 6 because I have Max. But it runs out tomorrow. In any case, It has been running smoothly. But I keep getting notifications that Opus 4.8 has had a number of critical errors.

What am I going to do after tomorrow? Just hope Opus 4.8 is good enough
 
Kimi K3 is competitive, but it only surpasses Fable 5 in some respects. Fable 5 is very good at maintaining a memory of all your sessions and connecting the dots.

Claude says:

"Thinking about comparing Claude's capabilities to Kimi K3 AI assistant


Let me check the latest on Kimi K3, since it's a recent release.


Timely question — K3 just launched yesterday (July 16), so your colleague is asking about something very fresh. Here's the honest picture.

"On capability, it's legitimately competitive. K3 is a 2.8-trillion-parameter model from Moonshot AI — the largest open-weight model ever announced — handling text, images, and video with a 1M-token context window. On the Artificial Analysis Intelligence Index it scores 57, placing fourth overall behind Claude Fable 5 (60), GPT-5.6 Sol (59), and Claude Opus 4.8 (56) — the first open-weight model to place that high. So roughly: it edges out Opus 4.8 on aggregate, trails Fable 5. Coding is its standout — it took the #1 spot on LMArena's Frontend Code Arena, placing first in six of seven frontend domains, ahead of Fable 5. On broader agentic knowledge work it beats Opus 4.8 but clearly trails Fable 5 (GDPval v2 Elo: 1,668 vs Fable's 1,760).

Three caveats to pass along. First, the benchmarks were run through different software harnesses depending on the test, which limits clean model-to-model conclusions — some gaps are within margin of error. Second, K3's hallucination rate increased compared to its predecessor, which matters more in valuation work than in frontend coding. Third, pricing: at $3/$15 per Mtok it's the same rate as Claude Sonnet 5 — no longer the ultra-cheap Chinese option, though still roughly a third of Fable 5's output cost.

For appraisers specifically, two considerations dominate over benchmarks. Sending client-confidential appraisal data to a Beijing-based company's API is a data-governance question most appraisal firms (and E&O carriers) would answer quickly, and not in K3's favor. But the flip side is genuinely interesting: the open weights are slated to arrive July 27, and a self-hostable frontier-class model is the one thing Claude can't offer at any price. For a confidentiality-bound pipeline like yours, a locally run K3 (or a distilled variant) could in principle solve the ZDR problem entirely — though serving a 2.8T model is a serious infrastructure lift, well beyond an M2 Ultra, until quantized or pruned versions appear.

Short version for your colleague: K3 is real competition at the Opus tier, best-in-class at frontend coding, behind Fable 5 on the hardest autonomous work — and the open-weights angle matters more for appraisal practice than the benchmark deltas do."

"Open weights" means the company publishes the actual trained parameters — the multi-terabyte file of numbers that isthe model — so anyone can download and run it on their own hardware. You already live this distinction: DINOv2 is open weights, which is why you can fine-tune it on your MLS corpus locally. The Claude API is the opposite pole.

Concretely, it buys you four things:

Data sovereignty. Inference happens on machines you control. No tokens leave your network, no retention policy to negotiate, no third party in the chain of custody. For your world this is the big one — it dissolves the ZDR problem, the E&O question, and the "confidential assignment results under USPAP" question in one move, because there's no counterparty at all.

Permanence. Once the weights are on your disk, nobody can deprecate the model, change its behavior, raise its price, or alter retention terms. For a Daubert-defensible methodology, that's not a small point — you can pin the exact model version used in an analysis and reproduce results years later, the same way you pin an R package version. API models get updated or retired under your feet.

Modifiability. You can fine-tune, quantize, distill, or LoRA-adapt the weights for your domain — say, adapting it on appraisal reports and USPAP text the way you're adapting DINOv2 for condition ratings. Closed APIs offer at most limited fine-tuning on the vendor's terms.

Price decoupling. You pay for compute, not tokens. Third-party hosts compete to serve open models, driving inference cost toward hardware cost, and for steady high-volume workloads self-hosting can undercut any API.

The honest fine print: open weights is not open source. You typically don't get the training data, the training code, or the data pipeline — so you can't audit what it learned or rebuild it, and licenses often carry use restrictions. And the capability tax is real: running a 2.8T-parameter model well takes a rack of H100s, not a Mac Studio, so in practice you'd either wait for distilled/quantized variants or use a hosting provider — which reintroduces a counterparty, though one you choose and contract with under your terms (including US-based hosts with zero-retention agreements)."
 
So, when Fireworks in Redwood City (CA) starts running Kimi K3, especially once I tie DINOv2 into my app to process photos, I will look into using it. In that case, Beijing won't get my info. -- And in fact, that is an issue if you are concerned about confidentiality, which should always concern appraisers.

I reviewed my usage and decided I don't need Claude Max at $200/month, so I'm dropping to $100/month and switching to "credits" to use Fable 5 when I need it. In fact, I might survive on Claude $20/month if I buy credits as needed for Fable 5 - which is a couple of times faster than Opus 4.8.
 
RCA said:
So, when Fireworks in Redwood City (CA) starts running Kimi K3, especially once I tie DINOv2 into my app to process photos, I will look into using it. In that case, Beijing won't get my info. -- And in fact, that is an issue if you are concerned about confidentiality, which should always concern appraisers.

I reviewed my usage and decided I don't need Claude Max at $200/month, so I'm dropping to $100/month and switching to "credits" to use Fable 5 when I need it. In fact, I might survive on Claude $20/month if I buy credits as needed for Fable 5 - which is a couple of times faster than Opus 4.8.
Click to expand...

K3 is the exact point you made about Macs being better at your AI interests then high-end Nvidia GPUs are for mine. But being able to vibe code as well as modern AI are I'm finding more usage out of less VRAM intensive tasks. And being able to do so much from your phone at 2:00 a.m. is literally keeping me up at night lol

Screenshot_20260718_095602_Digital Wellbeing.jpg
 
I asked one question in Claude and it resets in three freaking days. It's been good knowing you Fable 5

Screenshot_20260719_091412_Claude.jpg
 
Digger88 said:
I asked one question in Claude and it resets in three freaking days. It's been good knowing you Fable 5

View attachment 110553
Click to expand...

Claude often gets excited when I communicate with it. Ra!! Ra!! - It screams! And it says I use only 7% of my credits per week? It seems like it uses quite a bit of time on the tasks I give it, if you ask me. Maybe you are giving it too much boring yet demanding work, asking it to generate images all the time. I can tell you it must learn a lot from me, if it is learning anything, and it surely is at least on my system, but I believe it is not just local to me. Hard to say. I can only venture to guess. Also, ChatGPT and Grok seem very positive to me and know my history well.

So, if you have genuine experience and are rich with ideas, you will find that that is exactly what THEY are interested in. You just have to hope they keep your info confidential. If not, your best defense is to make everything as complicated as possible and hope they don't object. But of course you need to be able to manage any complexity you create.

Also. - You are your own worst enemy. But those whose livelihood you are perceived to threaten are also very potential enemies. The world is a VERY corrupt place. Honesty, what little exists in the world, is most often just a show - a facade - necessary for survival. Survival comes first over the good of society or mankind, or if you prefer, the World. We see that now in Ukraine, with the firing of Federov, who was a threat to the Old Military Foggies whom he was trying to push out. And Zelensky is no hero really. He will wind up with a reputation of corruption. That's why the WH doesn't really like him. He is the epitome of a Corrupt Politician. Only he is very smart, clever, and puts on a very good act.

Bezos, the Waltons, and most of the rich in the US are just looking out for themselves and, in some clans, for their families. The liberals are even worse, paltry corrupt officials with undying loyalty to their "Party".

Corruption everywhere, and in the highest circles. And even the MAGA and set back and consider the real constitution of Trump, who seems to be bipolar with respect to concern for the common good vs his good.

History will judge in the end.
As for appraisers, they are weak as a group. Very weak. The AMCs, lenders, GSEs, and Appraisal Organizations who are monitored directly or indirectly by the ASC who is monitored by the FFIEC, which pulls it members from the Federal Reserve Board:

www.federalreserve.gov

March 19, 2026 -- Notice of Open Board Meeting

The Federal Reserve Board of Governors in Washington DC.
www.federalreserve.gov

And these people mostly read from printed pages written by the unknown - may I submit - subversives who control our government. And, honestly, they probably know very little about appraisal. They just hand off unknown-sourced broad directives to the appraisal institutions for them to implement one way or another. These appraisal institutions, concerned 100% with satisfying their monitors, do then mostly what is expedient and least risky - and otherwise what is best for their pocketbooks and political power.

So, don't tell me about "monitoring" and "oversight." The only monitoring and oversight is by the most corrupt powers in lending, the AMCs and the GSEs - and the private corporations and appraisal organizations that they depend on. Odds are that anyone in appraisal with any say or any power who has been around a couple of decades is corrupt to the core.
 
You must log in or register to reply here.
Find a Real Estate Appraiser - Enter Zip Code

Copyright © 2000-, AppraisersForum.com, All Rights Reserved
AppraisersForum.com is proudly hosted by the folks at AppraiserSites.com
Back
Top