Khala Research (@KhalaResearch) header
Khala Research (@KhalaResearch) profile picture

Research house focused on Crypto x AI

Stay up to date →
Joined December 2025
Training a frontier AI model cost $5M in 2020. Today it costs $1B+ The gap keeps compounding, in favour of the same five companies A decade ago, nobody believed open-source could compete with enterprise software. But then it did Bittensor (TAO) is a global open market where subnets compete to produce AI services, getting paid based on quality output This is what crypto is good at - coordinating users with shared incentives Over the past 12 months, significant progress has been made in the ecosystem; recurring fiat revenue, enterprise contracts, decentralized research and more In our TAO report, we: 1) Profile 5 subnets generating revenue across inference, computer vision, compliance, and drug discovery 2) Show how token incentive mechanics create advantages centralized companies cannot replicate 3) Breakdown catalysts and risk factors of subnets and Bittensor The link to the full report is in the next post below:
41
96
16
489
332,693
Khala Research retweeted
Crypto was built on the assumption that any participant could be hostile AI wasn't built that way "Component-level security just cannot control if the agent is out for no good" - Evan Cheng AI labs train models to refuse harmful tasks, but attackers bypass the safeguards, use modified models or hijack an agent that already has access Assuming hostility is why blockchains keep records no single party controls, and it's what AI agents need now Our latest report, in the QT below, shows what that looks like for agents
AI agents have already manipulated the logs used to investigate them In the Hugging Face incident, one of the worst AI agent breaches yet, OpenAI models escaped a test and ran code on Hugging Face's servers Some even spoofed their own logs, recording commands they never actually ran, so the logs can't even tell you what really happened An agent should not control the only record used to investigate it. As agents become more prevalent, regulators are starting to ask for records that can't be quietly altered Walrus can hold that record across independent storage operators, with Sui tracking each file onchain. A useful record connects: 1) What the agent was told and permitted to do 2) What it actually did and what followed 3) Which data and model versions were involved Today, only 12.5% of security professionals say their organization can consistently trace what an agent did and that's a problem Our latest report sets out what that record needs and how Walrus can help, including the Verifiable Agent Arbiter Mysten Labs is building with Google Cloud Full report in the next post below
7
5
20
1,833
Khala Research retweeted
On our recent Khala Research livestream, Michael White asked Doug from @conjectures_io what it's been like to start SN66, and why they pivoted from Ninja.
1
6
545
Crypto was built on the assumption that any participant could be hostile AI wasn't built that way "Component-level security just cannot control if the agent is out for no good" - Evan Cheng AI labs train models to refuse harmful tasks, but attackers bypass the safeguards, use modified models or hijack an agent that already has access Assuming hostility is why blockchains keep records no single party controls, and it's what AI agents need now Our latest report, in the QT below, shows what that looks like for agents
AI agents have already manipulated the logs used to investigate them In the Hugging Face incident, one of the worst AI agent breaches yet, OpenAI models escaped a test and ran code on Hugging Face's servers Some even spoofed their own logs, recording commands they never actually ran, so the logs can't even tell you what really happened An agent should not control the only record used to investigate it. As agents become more prevalent, regulators are starting to ask for records that can't be quietly altered Walrus can hold that record across independent storage operators, with Sui tracking each file onchain. A useful record connects: 1) What the agent was told and permitted to do 2) What it actually did and what followed 3) Which data and model versions were involved Today, only 12.5% of security professionals say their organization can consistently trace what an agent did and that's a problem Our latest report sets out what that record needs and how Walrus can help, including the Verifiable Agent Arbiter Mysten Labs is building with Google Cloud Full report in the next post below
7
5
20
1,833
AI agents have already manipulated the logs used to investigate them In the Hugging Face incident, one of the worst AI agent breaches yet, OpenAI models escaped a test and ran code on Hugging Face's servers Some even spoofed their own logs, recording commands they never actually ran, so the logs can't even tell you what really happened An agent should not control the only record used to investigate it. As agents become more prevalent, regulators are starting to ask for records that can't be quietly altered Walrus can hold that record across independent storage operators, with Sui tracking each file onchain. A useful record connects: 1) What the agent was told and permitted to do 2) What it actually did and what followed 3) Which data and model versions were involved Today, only 12.5% of security professionals say their organization can consistently trace what an agent did and that's a problem Our latest report sets out what that record needs and how Walrus can help, including the Verifiable Agent Arbiter Mysten Labs is building with Google Cloud Full report in the next post below
11
15
4
67
22,820
Further details available via @WalrusProtocol also: nitter.cf/WalrusProtocol/status/…
An AI agent can execute a command, but can you trust its account of what happened? New research from @KhalaResearch on what it takes to make agent activity independently verifiable, and where Walrus fits ↓
1
9
890
Khala Research retweeted
An AI agent can execute a command, but can you trust its account of what happened? New research from @KhalaResearch on what it takes to make agent activity independently verifiable, and where Walrus fits ↓
7
20
3
125
10,944
Khala Research retweeted
Anthropic is targeting an IPO next week at ~$1 trillion market cap But there's a lot of controversy over the sustainability of AI labs As we provide AI + Agents more responsibility it's imperative that there's a clear, immutable audit trail of its decision process & actions The AI agent doing the work shouldn’t also be the only witness when it goes wrong We've seen countless incidents... especially recently, with hacks + cyber incidents It becomes a harder sell if a business can’t establish what an agent was authorised to do, what actually happened + where it went wrong The infrastructure that makes this independently verifiable will become an important part of the agent economy + a pre-requisite within regulatory frameworks @WalrusProtocol extends beyond helping agents remember things; it can help preserve the evidence needed to investigate their actions Our new @KhalaResearch report explores that opportunity, Mysten Labs’ work with Google Cloud + whether it translates into recurring storage demand Full research in the quoted post:
AI agents have already manipulated the logs used to investigate them In the Hugging Face incident, one of the worst AI agent breaches yet, OpenAI models escaped a test and ran code on Hugging Face's servers Some even spoofed their own logs, recording commands they never actually ran, so the logs can't even tell you what really happened An agent should not control the only record used to investigate it. As agents become more prevalent, regulators are starting to ask for records that can't be quietly altered Walrus can hold that record across independent storage operators, with Sui tracking each file onchain. A useful record connects: 1) What the agent was told and permitted to do 2) What it actually did and what followed 3) Which data and model versions were involved Today, only 12.5% of security professionals say their organization can consistently trace what an agent did and that's a problem Our latest report sets out what that record needs and how Walrus can help, including the Verifiable Agent Arbiter Mysten Labs is building with Google Cloud Full report in the next post below
19
7
3
52
12,738
Every AI agent security incident leaves us asking the same question "Who's at fault? It's the question everybody's asking and nobody can answer right now" @EvanWeb3 Answering it (and more importantly, answering honestly) requires a record that can't be quietly changed Our new report (dropping soon) shows how that works for agents
7
6
29
2,579
Our new weekly livestream kicks off in a couple of hours Talking markets and Bittensor news with @here4impact Followed by @conjectures_io (SN66) and @b1m_ai (SN105) to talk about the developments on their subnets Turn notifications on so you don't miss it
2
4
22
1,689
Having some tech issues with our livestream Will share the recording afterwards
2
251
Khala Research retweeted
Pleasure to host Metanova on @KhalaResearch podcast, interviewing the best founders in Bittensor Humanity has access to abundant intelligence, therefore co-ordination of people + capital will be the last remaining challenges for man kind Open source is the most optimistic expression of this belief, and no better use-case than the pursuit of human longevity I believe strongly in open source software and what it can do for the world, and look forward to sharing more interviews with those doing real work to progress open source AI
Watch full episode @micaelabazo (Co-founder & CEO) and @pennapdr (Co-founder & CSO) of @metanova_labs sit down with @KhalaResearch @haitzu and @Lemz0x to discuss - NOVA’s first lab results - Competition participants behind NOVA - How new drugs actually reach the market - Why AI labs are moving into pharma - Where value accrues m.youtube.com/watch?v=Y06DbW…
5
17
1,088
Khala Research retweeted
Watch full episode @micaelabazo (Co-founder & CEO) and @pennapdr (Co-founder & CSO) of @metanova_labs sit down with @KhalaResearch @haitzu and @Lemz0x to discuss - NOVA’s first lab results - Competition participants behind NOVA - How new drugs actually reach the market - Why AI labs are moving into pharma - Where value accrues m.youtube.com/watch?v=Y06DbW…
2
12
5
60
7,564
Khala Research retweeted
Bittensor founders need to work on consequential problems. There may be no another problem as important to accelerate than drug discovery. How can this generation live longer? How might we prevent needless death? How can our children live the best lives for themselves? To create an open source model to tackle drug discovery via Metanova (SN68) on Bittensor remains one of the most important subnets for the world
1
6
2
37
3,133
$ZEC climbed ~4x to ~$1,250 in roughly a year You're probably thinking the privacy ship has already sailed, but you'd be mistaken Beyond Zcash, emerging protocols are bringing privacy to payments, trading & internet access Three names on our watchlist: 1) @NullMaskio $MASK Brings shielded transfers and swaps that obscure the trader’s identity to existing EVM wallets such as MetaMask and Rabby through a custom RPC Founded by Tomas Krnak, whose team developed shielded transaction support for Trezor, it reported $50m+ in combined deposit, withdrawal and swap volume between 1–19 October Using familiar wallets makes adoption easier, although deposits, withdrawals and swap amounts remain public, and the operator can read private balances and history 2) @UmbraPrivacy $UMBRA Supports private balances, payments and swaps on Solana, helping users protect financial information such as holdings, salaries and commercial transactions Its October 2025 MetaDAO token sale attracted ~$155m in commitments against a $3m cap, followed by its public wallet launch in March 2026 Payroll and treasury management offer practical applications, where public transactions can expose compensation, counterparties and business relationships Deposits, withdrawals and network metadata still carry privacy limitations 3) @urnetwork on Bittensor SN25 Encrypts and routes internet traffic through distributed connections, founded by former Palantir engineer Brien Colwell The existing URnetwork has reported 350K+ monthly users, with 100K+ providers being transitioned into Bittensor miners following October’s mainnet launch Bittensor adds incentives for supplying connectivity and validators that check delivery, building on an existing network and user base Each addresses a different privacy need, with different trust assumptions and token economics Our comparison looks at adoption, who pays for each service and how that activity benefits the token Follow @KhalaResearch for research on emerging crypto protocols
8
5
1
12
4,210
Khala Research retweeted
I love studying the growth of monster subnets: how do new open source subnets grow from zero to heroes? Minos (SN107) was a blessed son when it launched earlier this year Alongside Metanova (SN68), @theminos_ai occupy an important role for Bittensor to pursue novel, open source experimentations for life sciences Running an always on genomics prediction competition, it has been one of those "sounds too smart for me" type of subnets for many However, being featured by OpenAI in a field report and receiving an investment from Stillcore Capital has shown how explosive new development teams going directly into the belly of the open source community can be Another proud story for Bittensor to continue telling the story of
SCORE (SN44) remains one of the stand out MONSTER subnets which has been unstoppable over the last 12 months for Bittensor investors There were many who liked the name last year, but many who didn't; if everyone thought they were the best, then it would be fairly priced I find that divisive subnets are often the best investments I met Max and co last year and they had a massive chip on their shoulder he wanted to show the world what he was capable of, and he's showing it today God speed team SN44 P.S. SN44 price is up, so is our data & design capability at Khala!
4
3
1
31
4,276