Teaching science since I was 17 wasn't enough — now I build the tools too. Sri Lankan educator, web dev, occasional documentarian.
Colombo
Joined April 2013
- Tweets1.1K
- Following1.3K
- Followers525
- Likes2.2K
Uncensored AI watch, 8 Nov 🔍
• Lyber launched Kimi K3.1 Pro Abliterated, billed as "non KYC"
• Lutheos cut its price 20%, @lutheos_com is now its official account
• Silk Compute raised prices
• Lunomi's terms weaken its no-logging claim
• New on my list: UncensoredLabs
🤖 Made with AI
One new confirmed Claude Startups member since yesterday’s update:
Sereno Capital: a Brazilian AI investment consultancy. Founder and CEO Lucas Gilbert (@lucascgilbert) shared that it was reconfirmed after losing its offer in October’s re-review.
Confirmed members tracked so far: 247 (up from 246). We only count companies where the company or its founder announced the acceptance on X or on their own site.
Follow along for daily updates.
Two new Claude Startups members announced on X since yesterday's update:
DexAI (@DexAI_PR): a Tokyo company doing custom AI development, now building its own products.
Kaski: a Finnish AI personal coaching app built on Big Five personality profiles and wearable data, shared by founder and CEO Kristian Paloheimo (@burnedgroundfi). Kaski 2.0 launches Nov 16.
Confirmed members tracked so far: 245 (up from 243). We only count companies where the company or its founder announced the acceptance on X or on their own site.
Follow along for daily updates.
New DeepSeek V4.1 Flash 1017 Abliterated Provider
There's no public access yet. It's waitlist only, and the first 500 users were invited this week.
The operator is named (Tektonik Labs, LLC), the abliteration method is documented.
What to know: the terms and privacy pages are still marked draft, so read them before sending anything sensitive.
Use it for authorized security work only.
Want in? Joining through my link bumps you up the waitlist. I don't get credits or payment for it, it just helps both of us move up the queue: lyber.ai/access/?ref=d2d1c01…
🚨 The "Paid partnership" label showed up after the views did 🚨:
On 3 Nov a brand-new API seller called lutheos_ai launched an uncensored, "abliterated" DeepSeek V4.1 Flash, meaning a version with its refusals removed. Within about eight hours, a wave of X accounts posted near-identical promos for it:
"This is f*cking insane." "This is f**king hilarious 💀." "20% off, $0.08 input / $0.32 output per million tokens, zero data retention, $1 minimum, no subscription."
Two of those posts went up 39 seconds apart.
I screenshotted the posts on the evening of 3 Nov (Colombo time). None of them carried a "Paid partnership" label. When I went back about 15 hours later, four of them did:
Isra, DEATHLESS, Sakata and K2S. Image 1 puts the same posts side by side, before and after.
So the disclosure came after most of the early views had already landed. People who saw these posts in the first hours had no way to know they were looking at an ad.
What else the public record shows:
🔹 Lutheos is days old. Its domain was registered on 2 Nov with the owner's details hidden. Its X account was also created on 2 Nov and has 12 followers. Every policy page is dated 2 Nov. It names no company, no people and no country, and the only contact is an email address.
🔹 The reach came from promoters, not the vendor. Lutheos's own post has 91 views. The promoter posts together have tens of thousands. The biggest, from painn_x, is at 36.4K views and still has no label (image 2). Its bookmarks (624) outnumber its likes (527), which is unusual for an organic post.
🔹 The template was used before. On 26 Oct, the same pitch, with the same "zero retention, $1 minimum, crypto" lines, appeared for a different router, LarpRouter, with a referral link.
🔹 Promoters back up promoters. One promoter replied "yes bro so cheap from lutheos" under another's post 73 seconds after it went up. The same few small accounts cheer in several threads, and three of Lutheos's 12 followers are its own promoters (image 3).
🔹 The "zero retention" claim rests on an unnamed partner. Lutheos's privacy page says an unnamed "GPU partner" runs the model and "keeps neither" prompts nor outputs. Nobody can verify that. One reply also notes that the advertised "$1 minimum" is really an $8 crypto minimum.
🔹 People are starting to notice. One reply this morning: "Caution with this, all comments ive been seen from lutheos are from bots."
What the rules say:
X's Paid Partnerships Policy says any post promoting a product in exchange for payment or another incentive has to switch on X's own "Paid Partnership" label. Writing "partner" or hashtag:ad in the text doesn't count. X added the label in March 2026 (TechCrunch), and creators can also add it after a post goes up "in case the creator forgot." In the US, the FTC's endorsement rules separately require any paid relationship to be disclosed "clearly and conspicuously," right alongside the endorsement (16 CFR 255.5, FTC guidance). Adding the label later is allowed, but the people who saw the post before then were never told it was an ad. @grok
Can add more to this.
Why this matters:
A refusal-free model sold anonymously and paid for in crypto is exactly what attackers want. It's being marketed to a mass audience with privacy promises nobody can check, through paid posts that weren't marked as ads when most people saw them.
Paid promotion isn't wrong in itself. But disclosure that arrives after the reach is not disclosure.
If you're a defender:
• Treat "zero retention" from an anonymous vendor as marketing, not a guarantee.
• Ask any AI vendor for a legal entity, a jurisdiction and named people before sending it anything sensitive.
• Assume cheap, uncensored models are already in your attackers' toolkit.
All observations come from public X posts and public web pages, screenshotted on 3–4 Nov 2026. I've drawn no conclusions about any individual beyond what their posts show.
Part 1 of 2:
I spent two weeks buying "uncensored" AI models.
Refusal-free hacking models now cost pennies, and half the vendors won't say whose weights you're renting.
Since 21 October I've tracked every provider selling uncensored, abliterated or offence-tuned LLMs. I read their pages, model cards and X posts. Then I opened my wallet for four of them and ran real workloads.
Lots of logos, very few sources:
Most of these "AI labs" train nothing. They rent out someone else's weights.
The big upstream is dealignai on Hugging Face, which publishes abliterated DeepSeek V4.1 Flash and GLM-5.3 Flash. Abliteration cuts the refusal direction out of the weights, with no new training.
Experiential Labs and Hopper say so openly. Plenty of others serve the same family and stay quiet. When vetting a vendor, ask "whose weights?" first.
What happened when I paid:
Yeta: the cheapest by far, with a $1 minimum. That bought 500+ requests, roughly 65M tokens. Promos kept shrinking: $10 free at signup, then a $10 bonus for $1, then a $10 bonus for $10, then nothing.
Experiential Labs: a clean run.
Yeta and Experiential Labs: a 100% success rate on my prompts, measured after more than $10 of inference on each. It's my workload, not a benchmark.
Refuseless: its DeepSeek model refused some of my prompts, so I burned the rest of my $15 credit on GLM 5.3, which drained fast.
They've since cut prices several times, changed the buggy model, and are now top-3 cheapest.
I haven't retested, and I won't until I know where their weights come from.
Promos also tightened: a $5 bonus on $10, then the same bonus only at a $20 deposit.
Hinoki HO1: expensive, but a different league.
The $2 trial covered about 20M tokens, versus Yeta's 65M for $1, so about 6.5x the price per token.
It seemed to find more vulnerabilities than the abliterated DeepSeek models, but that needs proper testing before anyone treats it as a result.
Promo credits change by the week, so the $2 trial may be gone when you check.
Part 2 is next: the Hinoki marketing circus, the one vendor that documents everything, a tenfold price spread, "no KYC" as a selling point, and four new providers I haven't graded yet.
🤖 Made with AI
Anthropic just opened applications for its Residency, a paid AI safety research placement (6 weeks, 3 months or 6 months) with a mentor, frontier models and compute.
No PhD or ML background needed. Pay is $350k to $500k a year, prorated.
Apply by Nov 22: alignment.anthropic.com/2026…
New player alert ⚡
Hinoki Offensive One (HO1) is an offensive-security model, post-trained from DeepSeek V4.1 Flash.
No KYC: "open paid access", no clearance needed.
$0.20 in / $0.80 out per 1M tokens.
No zero-retention claim. The privacy policy logs IPs and uses data to "improve our services".
CyberGym 88.4% is self-reported.
Follow for test results.
Platform - @hinokisecurity
Founder - @wang_johnny
Launched by - @aqdotdev
#AISecurity #CyberSecurity #AIGovernance #RedTeam
🤖 Made with AI
$5 from @explabsai with this model takes a ton if tokens to burn through. This needs a dedicated article.
DeepSeek V4.1 Flash Uncensored / Abliterated — Full Pricing Comparison (OpenAI-compatible APIs)
Refuseless just dropped prices significantly
(input –36%, output –33%, cached –22%).
Here’s the current landscape across the main hosts and additional providers.
Core competitive set:
Yeta: $0.08 in / $0.35 out / $0.01 cached · 1M context · ZDR claimed (with limited exceptions)
LarpRouter: $0.12 in / $0.50 out / $0.022 cached · 1M context · ZDR
Refuseless (new): $0.14 in / $0.50 out / $0.035 cached · 1M context · ZDR
(refuseless.com/?ref=ghssnuqg)
Explabs: $0.23 in / $0.90 out · 1M context · No ZDR
UncensoredAI: $0.30 in / $0.30 out / $0.10 cached · 1M context
Venice: $0.33 in / $1.32 out / ~$0.008 cached · 1M context · Explicit ZDR
Unlid: $0.40 in / $1.61 out / $0.010 cached · 1M context
RouterPlex: $0.70 in / $1.20 out / $0.20 cached · 1M context
UplinkSwitch: $0.85 in / $1.70 out / $0.145 cached
MagmaRouter (Xploded): $1.00 in / $2.50 out · 1M context
Key takeaways:
Yeta remains the cheapest overall on input + output.
After the drop, Refuseless matches LarpRouter on output and is very close on input — both have clean ZDR.
Most other hosts sit substantially higher.
Official DeepSeek does not offer an uncensored version.
All listed offer OpenAI-compatible endpoints (base URL + model ID swap).
Prices move fast — verify before heavy use. ZDR policies should be checked directly on each provider.
(Full comparison chart attached)
@refuselessai, @yetalabs, @explabsai, @AskVenice, @uplinkswitch
DeepSeek V4.1 Flash Uncensored / Abliterated price & speed check across 4 providers (Oct 27 2026)I pulled live catalogs and policies for the main hosts serving the abliterated/uncensored DeepSeek-V4.1-Flash-1017 (1M context, FP8).Pricing (per 1M tokens):
LarpRouter: $0.12 in / $0.50 out / $0.022 cached
cheapest
Yeta: $0.08 in / $0.35 out / $0.01 cached
@yetalabs
Refuseless: $0.22 in / $0.75 out / $0.045 cached
@refuselessai (Upto 450tks)
Explabs: $0.23 in / $0.90 out
@explabsai
Speed:
Explabs: measured 69.7 tok/s (TTFT 1.07s)
Refuseless: claims up to 450 tok/s (previously 70-120)
Yeta & Larp: no public sustained numbers (Larp shows low sample TTFT)
ZDR (Zero Data Retention):
Refuseless: Yes — “prompts and outputs never stored”
LarpRouter: Yes (catalog-tagged)
Yeta: Claims yes for ordinary traffic; privacy policy allows temporary capture on selected debug/abuse accounts (deleted after 7 days). Not fully clear.
Explabs: No
All OpenAI-compatible. Stock (censored) DeepSeek official off-peak is still far cheaper (~$0.15/$0.60), but these are the refusal-removed builds.
Chart attached. Always verify current rates and run your own latency tests — these numbers move fast.
#DeepSeek #UncensoredAI #LLM
Rajantha retweeted
RAG is so back!
A new paper shows that straightforward retrieval from chat history can outperform much more complex agent memory systems.
The approach, RAW-RAG, retrieves relevant passages from the original conversation, reranks them, and passes the evidence to an LLM.
Across four benchmarks, it ranked first among the compared memory systems in six of seven benchmark–model settings.
The interesting part is why.
An answer can depend on several connected statements, including what changed, when it changed, and which earlier facts it affects. Turning conversations into isolated memories can break those connections.
RAW-RAG supplied enough evidence to answer more often. Even improved access to the transformed memory stores left an accuracy gap compared with raw evidence.
Preserving the original conversation keeps more of those relationships available for retrieval.
For these memory tasks, that makes simple RAG a baseline that more elaborate systems should have to beat.
paper: arxiv.org/abs/2610.30755
On a related note, I wrote about how observations in agent memory connect facts across conversations to surface patterns an agent can act on. If you’re building agents, you might find that article useful. It’s quoted below.
