I count what others feel | Prediction market portfolios | Sport·Math·Edge | Models, not magic | @CoinMarketCap https://xitter.cf/t.co/krp0rc4MKl
Joined May 2012
- Tweets15.1K
- Following751
- Followers2.4K
- Likes52K
ghøst retweeted
67¢ on Polymarket doesn't necessarily mean you're paying 67¢
I checked a real tennis market. After taker fees, the effective price was 68.1¢
MagicMarkets showed 64.5¢ for the same outcome. Here's what that meant for a $100 trade
I compared Michele Ribecai vs Jacopo Vasami using Thumper
For Michele Ribecai:
-> Polymarket payout: $147
-> MagicMarkets payout: $155
-> $8 more on MagicMarkets
Same match. Same bet size. Different prices
Thumper is a free Chrome extension by @MagicMarkets that shows MagicMarkets prices alongside Polymarket
MagicMarkets offers zero commission, but isn't always cheaper
Don't just check the displayed odds. Check the actual payout
1/ i built a chrome extension for @magicmarkets that shows you when polymarket isn't the cheapest place to take a sports position
it's called thumper. Free to use, no account required 🧵
ghøst retweeted
Insurance claims run on photos. So what happens when the photo is lying?
I built my #PerturbVideo submission around that question, using Angle 3: The claim that never happened
A real example was reported by The Guardian in May 2024. A van photo was taken from a tradesman's social media, edited to show a cracked bumper, and sent to UK insurer LV= with a fake invoice for more than £1,000. LV= had concerns and passed it to its fraud team. Allianz UK said claims involving app-edited images, videos and documents rose 300% between 2021-22 and 2022-23
My video then plays out a clearly labeled, made-up scenario where an automated claims model approves a doctored photo. A few tuned pixels that a person would never notice can leave a model wrong and still confident
The cost doesn't just vanish. Insurers cover the payout, and honest policyholders feel it in their premiums. The Coalition Against Insurance Fraud estimates US insurance fraud at $308.6 billion a year in its 2022 study, which is an industry figure and not something Perturb measured
Perturb @perturbaix flips the usual order and breaks the model first, before a fraudster gets the chance. On Bittensor Subnet 26, miners search for the smallest change that makes a model's answer flip. Validators confirm the flip is real and the change stays invisible, copied work gets no reward, and whoever finds the best attacks earns the most $TAO
For the team behind the model, that testing turns into a robustness score, a vulnerability heatmap, verified attack data to retrain on and an on-chain certificate of testing. Details live at perturbai.io
As of now, image-model testing is live and the network has been running since May 2026. Text, LLM and multimodal attacks are still on the roadmap, and November presales are a target rather than a done deal
ghøst retweeted
The price on screen isnt always the price you pay
On Liverpool vs Villarreal, Polymarkets draw costs 21.8c all-in, while MagicMarkets offers 20.8c
Thats a $12 difference in potential payout on a $100 trade
Across 100 comparable trades, thats $1,200. Same stake, different execution price
Thumper is a free Chrome extension that puts @magicmarkets prices next to Polymarket sports markets, so you can compare the real cost before trading
Check the price
Then take the trade
1/ i built a chrome extension for @magicmarkets that shows you when polymarket isn't the cheapest place to take a sports position
it's called thumper. Free to use, no account required 🧵
ghøst retweeted
Paying a global network to break AI models and generate retraining data beats a closed QA team every time. SN26 is one of the cleanest real-world use cases on Bittensor / bittensor:native
Angle 1: The Taped Sign
A tiny piece of tape
A very different reading
In 2020, McAfee showed that tape could make an older vision system read 35 mph as 85 mph
For automakers, a misread sign can become a safety and liability risk. @perturbaix helps find these weaknesses before deployment
On Bittensor Subnet 26, miners compete to find tiny changes that fool AI models
Validators check that each attack works and remains imperceptible. Copies earn nothing
Customers get robustness scores, vulnerability heatmaps, and verified data for retraining
Image-model testing is live
Text, LLM, and multimodal attacks are still on the roadmap.
Find weaknesses before attackers do
#PerturbVideo
ghøst retweeted
AlphaRidge puts research, screening, and automation in one place instead of five tabs and three subscriptions.
A trader finds a setup on the screener, turns it into a rule, and the rule keeps running and alerting even with the browser closed.
@alpharidge_ai maps a few ways the workflow actually closes:
> Screening on news: Sort the whole market by 7-day sentiment, event clusters, and impact scores instead of just price and volume.
> From idea to rule: Paste a TradingView Pine script into RidgeScript, test it, then arm alerts to Discord, Telegram, or email.
> AI that shows its work: The analyst writes screens and marks up charts, then surfaces the exact platform data it used so you can check it.
The neat part is the news layer. Miners on SN45 turn hundreds of thousands of articles a day into screenable sentiment and events, so the screener can look at what is being said, not just what already happened.
Paper trading is free today. Live brokerage execution is still rolling out. The AI cites its sources, but you still have to confirm before anything is placed.
ghøst retweeted
I saw that @heydittoai walked through Ditto v2 at Exploit Summit, including the bit where an agent switches platforms and keeps the context.
Their own write-up uses a case that is basically one sentence: point the new agent at the old one and tell it to mind meld. No rebuild.
To give you some more context on this, a normal chat remembers you only inside that app. Claude does not know what you told Cursor, and a new agent starts from zero.
But if the spec changed in Telegram, or you moved from one coding agent to another, Ditto v2 keeps one memory under the tools and waits for a yes before it sends or edits anything risky.
So the flow would look something like this:
▸ You say it once, in Claude, Cursor, Codex or Grok, and the others can already use it.
▸ Telegram and Slack sit next to the docs and the code, so the agent can search the decision, not just the repo.
▸ A workflow or browser agent runs the job, and the step that could do damage waits for your approval.
Depending on the tool, you stay in that chat or leave with the memory exported.
This ties back to what I wrote before: the models got good at answering, but you are still the switchboard between them.
That’s the part Ditto v2 is building for, even when the handoff is a single sentence.
What other agent handoffs should I look into?
ghøst retweeted
SayGM (SN28) might be the best AI altcoin that noone is talking about 💎
This $TAO subnet has rapidly become a blue chip in the Bittensor ecoystem
Why?
First of all, their new desktop app allows anyone to access Claude, ChatGPT, Gemini, and Llama for up to 70% cheaper and privately
Currently at $5M ARR and growing 🤑
Secondly, their new growth marketing affiliate program "Proof of Customer affiliate pool" is turning heads:
- @say_gm_ pays people who refer new developers that actually spend money on its AI gateway
- You can earn up to $1 worth of SayGM Subnet 28 alpha for every $1 your referrals spend during their first 30 days
Example: you refer a developer who spends $1,000 -> you could earn up to $1,000 worth of SN28 alpha
- The rewards do not come from SayGM's normal business revenue or marketing budget. They come from a pool funded by Subnet 28 emissions
So now they have an amazing product and an effective growth strategy
Sky is the limit here 🚀
ghøst retweeted
One of the biggest announcements from @ExploitSummit Montreal may have flown under the radar.
@green_compute_ $SN110 announced another $4M data center expansion after already bringing a $3M facility online.
Their target? Roughly $300K/month in revenue from renewable-powered compute.
This comes after Green Compute reported $82K+ in profitable revenue in its FIRST month on Bittensor.
Real GPUs / Real enterprise customers / Real revenue. AND 10% of contract revenue flowing toward alpha buybacks.
If you missed them present at Exploit, here's the link to the full presentation:
xitter.cf/green_compute_/status/…
For those who were not able to join us, here`s our full presentation from Exploit Summit 2026 @ExploitSummit where we proudly announced our next big step. 🚀
ghøst retweeted
TAO wallets are becoming portfolio tools, not just key storage.
@CrucibleLabs is building the place where a Bittensor position should stay usable.
The wallet can hold TAO and alpha in one balance, swap across subnets, claim rewards, and route root yield through Smart Allocator.
The goal is keep custody with the user, cut the spreadsheet, and put earned rewards back to work.
Over time, the same surface expands into:
▸ Subnet-level performance
▸ Ledger on mobile
▸ Multi-address control
That gives the wallet a bigger role than sending and receiving.
It can support swaps, staking visibility, and reward allocation without a separate tool for each step.
The wallet becomes the balance sheet for a TAO position.
ghøst retweeted
imo @brandtensorAI is attacking subnet growth from too many angles at once.
i'm really curious how one team keeps brand, proof, community, and outside demand in the same system.
skip the agency-brochure side for a sec and look at what they actually force a subnet team to ship:
+ a clear value prop before the launch thread even exists
+ different messaging for miners and validators vs AI users and enterprises
+ a proof system so the strongest claims have evidence behind them
+ use-case and demo plans instead of another architecture dump
+ distribution that does not stop at the TAO timeline
for founders there's also a shared growth plan: what to say, where to say it, and how to measure if it worked
this is the brandtensor playbook: make the tech choosable, then take it to market
➜ the goal is simple: turn technical progress into trust, adoption, and people who actually use the product
respect to the team for going this hard on a problem most subnet founders treat as an afterthought.
ghøst retweeted
SN80 OpenRoboto is breaking one of robotics’ biggest walls: data locked inside closed companies and single ecosystems.
Shift turns real-world human work into robot training data, while @openroboto is distributing that network across Base and Robinhood Chain.
Physical AI, opened up. bittensor:native robinhood:0x6f63d869011f95274498023b4abfc00b30c34378
ghøst retweeted
sports traders know how important it is to be able to adjust a position mid-game.
that’s one of my favourite part of @magicmarkets.
you can trade tennis and soccer in-play, with sportsbook and prediction market liquidity aggregated into one book and 0% fees.
that's good tek.
🤝 Paid partnership
ghøst retweeted
This is a much bigger deal than “fast swaps.”
@taodotcom just made Bittensor subnet liquidity way more accessible.
USDC → dTAO in seconds.
dTAO → USDC in seconds.
No clunky route back through the ecosystem.
Every subnet just got a cleaner bridge to outside capital.
That’s how you turn subnet tokens from a niche crypto primitive into a real market.
bittensor:native
USDC to dTAO Crosschain swaps.
Now LIVE.
Swaps settle in seconds, it's FAST.
Like, seriously quick.
Directly from Subnets into USDC (Or ETH) and vice versa.
Try it today at TAO.com
ghøst retweeted
➥ @conjectures_io just published the cleanest version of the mechanism: pay for a machine-checked result on a problem that still has no answer key.
previously, most AI math sat on benchmarks. your compute recovered a keyed answer, a panel or a score decided if it “counted,” and a confident writeup could look finished while remaining unchecked.
the new flow is cleaner:
→ pick an open problem in the catalog
→ write a Lean proof or a counterexample with any model you want
→ pay a flat fee for one check against the pinned statement
→ the kernel returns yes or no
→ kernel pass is necessary, not sufficient; review still sits on gamed files and lifted proofs
so you keep the method private while quoting a public target, and only get paid once the file actually settles.
the practical effect is that discovery becomes payable. finding the argument can take months. checking it takes seconds. that gap is why strangers can be funded without trusting their story.
it’s a small surface change - one Lean file - but it removes a friction most research prizes still carry, by making the first verdict closer to how a proof assistant already thinks about truth.
open one problem at conjectures.io and read the statement.
ghøst retweeted
A major cost layer for coding agents has just made an extraordinary move on the savings print.
@somasubnet, currently live inside GitHub Copilot on DeepSeek V4 Pro, has just shipped a new compression algorithm that improves the previous one by 5pp - so you can now save up to 15%.
With this update, SOMA is focusing on 2 pillars in the next run:
• putting the stronger compressor to work in the same Copilot loop developers already use, instead of asking them to swap models
• pushing the savings ceiling from the ~10% launch print toward the 20% target the team has been aiming at
In case you're into agent workflows and a Copilot user, @somasubnet should be the compression layer for cutting repeated context before it hits the model.
#Bittensor #TAO #Somasubnet
We taught SOMA a new trick.
The new compression algorithm improved on the previous one by 5pp, so you can now save up to 15%.
Put it to work: app.thesoma.ai/register
Compress. Pay less.
ghøst retweeted
A Polymarket tennis trader up $2.6M on sports markets is now available to copy
I found a pretty interesting way to follow his trades without watching Polymarket 24/7
You can customize the copytrading strategy based on your own risk and position size
I checked the leaderboard on @arespro mobile and the performance metrics are super transparent
You can easily see which traders are performing well and choose who you want to follow
Great work by @morganlai on bringing this to mobile!
Copy trading in prediction markets is broken: slow fills, slippage, hard to find good wallets.
Ares Mobile Copy Trading for @Polymarket fixes all three.
Fastest execution. Lowest slippage. Curated wallets.
Like + RT + drop your Ares profile → 20 winners, $25 each. 9/16 🧵
ghøst retweeted
SABALENKA. PEGULA. GAUFF. RYBAKINA.
Tonight the US Open women’s final is priced in cents.
Four names. Top 4 seeds. First time since 1975 all of them are in the semis.
Match 1
Aryna Sabalenka vs Jessica Pegula
SAB 64.9c / 63.2c
PEG 36.7c / 35.4c
Volume already $43k
Two-time US Open champ vs the home crowd on Ashe.
64.9c is not a vibe. That is the market saying Sabalenka is about a 65% favorite.
Match 2
Coco Gauff vs Elena Rybakina
GAU 50.0c / 48.1c
RYB 51.8c / 50.4c
Volume $73k
Coin flip. Bigger book than the first semi.
Gauff on her court. Rybakina with the serve that travels at night.
That is not a bookie line.
That is a live price you can trade.
On @magicmarkets every cent is a probability.
65c = the market giving Sabalenka a ~65% chance.
50c / 52c on Gauff-Rybakina = the book cannot pick a winner yet.
Think 65c is too rich? Sell Sabalenka. Buy Pegula.
Think 50c on Gauff at home is cheap? Take her and exit when the break hits.
You do not wait for the last point.
Break of serve hits, the book reprices, you close and lock it in.
That is the difference.
MagicMarkets pulls liquidity from 25+ sportsbooks, exchanges and prediction markets into one book.
So when the night session opens you can still get filled. Pre-match. Live. And on the way out.
Zero commission. No winner limits.
Same book tonight. Two women’s semis. One men’s semi already listed for Friday:
Zverev vs Khachanov
Don’t pick a winner.
Trade where the market is wrong.
This is another US Open match I’m paying close attention to today
Aryna Sabalenka vs Jessica Pegula - I’m taking Sabalenka here
Pegula has been playing some really good tennis this season and I expect another competitive match
But I still think Sabalenka has the edge
She leads the H2H 9-4 and has beaten Pegula in their last two US Open meetings
Sabalenka’s power is a huge factor for me
If she can keep the unforced errors under control, I think she has enough firepower to take control of the match
The live markets on @magicmarkets:
→ Sabalenka to win - 64.9¢
→ Pegula to win - 36.7¢
And the game handicap:
→ Sabalenka -3 - 50.7¢
→ Pegula +3 - 52.3¢
I’m backing Sabalenka here
At around 65¢ on the winner market, I’m comfortable taking her side
She has already shown she can handle Pegula on this stage, and I’m expecting another big performance tonight
Who are you taking?
ghøst retweeted
SOMA is live for GitHub Copilot
It's a lightweight context compression layer, building to capture developers running coding agents who want lower token bills without changing models or rewriting their workflow.
@somasubnet differentiates others by being thinner:
- No new model to swap in
- Just compression + execution in front of DeepSeek V4 Pro
- The user remains inside Copilot, on the same agent loop
→ Looks more like an inference-cost layer > a typical AI-agent launch
Worth a look if you've been watching Copilot sessions eat context. Early Access credits and no platform fees are a nice touch too.
#Bittensor #Tao #Soma
SOMA is live for GitHub Copilot.
All Early Access emails have now been sent. If you signed up, check your inbox.
You can plug SOMA into Copilot and use DeepSeek V4 Pro with 10% savings. Every Early Access user gets $5 in credits to test it on real coding tasks.
There are no SOMA platform fees during Early Access.
If you missed Early Access, we’ll open SOMA to everyone soon.
ghøst retweeted
Robotics is getting smarter, but the real challenge is making robots more useful.
@openroboto is building an open robotics ecosystem where robots can learn, share skills and improve through AI.
If robots can learn from each other instead of starting from scratch, we could see robotics improve a lot faster. The next phase of robotics could look very different
