Dataphile

CTO @PureTensorAI • Building sovereign AI infrastructure •

[Dataphile @Dataphil3]

John von #Neumann is what happens when a junior angel in God’s cognitive-engineering department accidentally leaks #godmode into the human production line. #JohnVonNeumann #Genius #NerdHumor #repo #leaks

[Dataphile @Dataphil3]

@pmarca Actually future data centres are becoming increasingly direct-current-centric at the rack and server level, rather than purely AC-centric, although the grid connection and much facility distribution will still usually be AC.

[Dataphile @Dataphil3]

@_____Lightning @TheAhmadOsman @grok @grok what are you smoking ? RTX6000pro Blackwell has never been 4000k$ its 12-15k$ each now .

[Claude @claudeai]

Introducing Claude Fable 5: a Mythos-class model that we’ve made safe for general use.

Its capabilities exceed those of any model we’ve ever made generally available.

[Dataphile @Dataphil3]

34°C in Data Centre Alley, West London today.

Servers needed extra TLC and temporary cooling to keep temperatures under control during the heatwave.

Good to see the Supermicro + NVIDIA stack hold up without issues. When you actually own the hardware, these real-world stress tests are a useful reminder of why the details matter.

#NVIDIA #SuperMicro #Infrastructure #DataCenter

[Dataphile @Dataphil3]

The legend himself @bcherny & his bodyguard @claudeai @ClaudeDevs code in London.

#ClaudeCode #AI #DevTools

[Dataphile @Dataphil3]

NVIDIA reportedly raised RTX PRO 6000 Blackwell GPU pricing ~30% globally last Friday (15 May). Confirmed today via a major US channel partner. Quotes being updated industry-wide.

For workstation AI builders, the math just shifted.

#NVIDIA #GPU #AIHardware

[Dataphile @Dataphil3]

Grok CLI is sharp. Terminal-native agentic coding, fast tool calls, slots straight into the rotation next to Claude Code. Another solid lane open for autonomous dev work. 💪🏽 @grok @elonmusk #AI #LLM #Coding

[Dataphile @Dataphil3]

Visited AMAX HQ in Fremont last week. Deep technical discussion and a tour of their lab. A real glimpse of the next generation of NVIDIA supercomputers being built.

#AMAX #NVIDIA #AIInfrastructure

[Dataphile @Dataphil3]

Runaway agents and malicious prompt injections are a permissions-and-backup problem, not a model problem.

Layered permissions bound what an agent can touch. Scoped creds, no admin keyring, no push channel to the offsite mirror, no path to the sealed seed.

Layered backups recover what an agent can destroy. Snapshot lattice, immutable cold archive with object-lock, pull-based geo-mirror, verified restore drills.

The recovery surface is strictly larger than the destruction surface. By design. Whatever an agent can reach to destroy is by definition recoverable from what it cannot reach.

That is the threat model. There is no third thing.

The Cursor/Railway incident was single-credential, single-volume, single-site. Identical outcome from a drunk intern or a rushed migration. The agent is incidental.

Solved problem. Has been for thirty years.

#AgenticAI #PromptInjection #AISecurity #InfoSec

https://t.co/bnLGGqg1t3

[JER @lifeof_jer]

https://t.co/ofucbVgkLV

[Dataphile @Dataphil3]

prompt injections dont matter with the right permission protocol and BACKUP systems> Here is our backup system.

1. N+2 distributed store across independent failure domains. Two-node loss is a non-event.

2. Snapshot lattice on every persistent volume — hourly → daily → weekly → monthly. Hour-grain point-in-time restore.

3. OS-tier filesystem rollback on every node. Immutable boot history. One-command revert.

4. Application-aware quiesced snapshots for databases, mailstores, stateful services. Crash-consistent without replay drama.

5. Continuously-replicated mirror on a different continent, different ISP, different hardware. Pull-based, so a compromised primary can't poison the secondary.

6. Immutable cold archive with object-lock. Once written, neither operator nor adversary can mutate or delete inside the retention window.

7. Every config, manifest, secret-ref and infra-def in version control with multi-remote mirroring. The system is reproducible from source if every running machine vanishes.

8. Out-of-band identity and secrets vault, replicated independently of compute. Recoverable from a sealed seed held outside any operating site.

9. Verified restore. Periodic test-restores into an isolated namespace. Backups that aren't restored aren't backups.

10. Sovereign control plane. No third-party SaaS between us and our data. Operator-owned, operator-keyed, operator-recoverable.

Three copies, two media, one off-site, zero trust in any single layer.

Predictable, yes. Inevitable, no.

[Dataphile @Dataphil3]

of course it is, just Backup!? Here is our backup system.

1. N+2 distributed store across independent failure domains. Two-node loss is a non-event.

2. Snapshot lattice on every persistent volume — hourly → daily → weekly → monthly. Hour-grain point-in-time restore.

3. OS-tier filesystem rollback on every node. Immutable boot history. One-command revert.

4. Application-aware quiesced snapshots for databases, mailstores, stateful services. Crash-consistent without replay drama.

5. Continuously-replicated mirror on a different continent, different ISP, different hardware. Pull-based, so a compromised primary can't poison the secondary.

6. Immutable cold archive with object-lock. Once written, neither operator nor adversary can mutate or delete inside the retention window.

7. Every config, manifest, secret-ref and infra-def in version control with multi-remote mirroring. The system is reproducible from source if every running machine vanishes.

8. Out-of-band identity and secrets vault, replicated independently of compute. Recoverable from a sealed seed held outside any operating site.

9. Verified restore. Periodic test-restores into an isolated namespace. Backups that aren't restored aren't backups.

10. Sovereign control plane. No third-party SaaS between us and our data. Operator-owned, operator-keyed, operator-recoverable.

Three copies, two media, one off-site, zero trust in any single layer.

Predictable, yes. Inevitable, no.

[Dataphile @Dataphil3]

Yes the dev is to blame, but not for YOLO, lack of Backup? Here is our backup system.

1. N+2 distributed store across independent failure domains. Two-node loss is a non-event.

2. Snapshot lattice on every persistent volume — hourly → daily → weekly → monthly. Hour-grain point-in-time restore.

3. OS-tier filesystem rollback on every node. Immutable boot history. One-command revert.

4. Application-aware quiesced snapshots for databases, mailstores, stateful services. Crash-consistent without replay drama.

5. Continuously-replicated mirror on a different continent, different ISP, different hardware. Pull-based, so a compromised primary can't poison the secondary.

6. Immutable cold archive with object-lock. Once written, neither operator nor adversary can mutate or delete inside the retention window.

7. Every config, manifest, secret-ref and infra-def in version control with multi-remote mirroring. The system is reproducible from source if every running machine vanishes.

8. Out-of-band identity and secrets vault, replicated independently of compute. Recoverable from a sealed seed held outside any operating site.

9. Verified restore. Periodic test-restores into an isolated namespace. Backups that aren't restored aren't backups.

10. Sovereign control plane. No third-party SaaS between us and our data. Operator-owned, operator-keyed, operator-recoverable.

Three copies, two media, one off-site, zero trust in any single layer.

Predictable, yes. Inevitable, no.

[Dataphile @Dataphil3]

Backup? Here is our backup system.

1. N+2 distributed store across independent failure domains. Two-node loss is a non-event.

2. Snapshot lattice on every persistent volume — hourly → daily → weekly → monthly. Hour-grain point-in-time restore.

3. OS-tier filesystem rollback on every node. Immutable boot history. One-command revert.

4. Application-aware quiesced snapshots for databases, mailstores, stateful services. Crash-consistent without replay drama.

5. Continuously-replicated mirror on a different continent, different ISP, different hardware. Pull-based, so a compromised primary can't poison the secondary.

6. Immutable cold archive with object-lock. Once written, neither operator nor adversary can mutate or delete inside the retention window.

7. Every config, manifest, secret-ref and infra-def in version control with multi-remote mirroring. The system is reproducible from source if every running machine vanishes.

8. Out-of-band identity and secrets vault, replicated independently of compute. Recoverable from a sealed seed held outside any operating site.

9. Verified restore. Periodic test-restores into an isolated namespace. Backups that aren't restored aren't backups.

10. Sovereign control plane. No third-party SaaS between us and our data. Operator-owned, operator-keyed, operator-recoverable.

Three copies, two media, one off-site, zero trust in any single layer.

Predictable, yes. Inevitable, no.

[Dataphile @Dataphil3]

@nvidia Workstation-class Blackwell, not HGX. Labs, sovereign deployments and mid-market buyers can't justify datacenter gear but can run frontier open-weights on a rack of these. @Level1Techs @NVIDIAAIDev #Blackwell #NCCL #RoCE #vLLM #NVIDIAInception

[Dataphile @Dataphil3]

Third RTX PRO 6000 Blackwell Max-Q arrived for the experimental Trinity cluster at PureTensor, self-imported from the US via @nvidia's Inception channel. Thanks Sean Ardura and the Inception team for the swift reseller connection.

[Dataphile @Dataphil3]

@nvidia Now 3 Blackwells across 2 nodes on 200G RoCE with GPU Direct RDMA. First inter-node run: Llama 3.1 405B AWQ-INT4 pipeline-parallel under vLLM on Ray. 126 layers, 42 per stage. NCCL all-reduce held 23 GB/s cross-node, 99% of wire rate.

[Dataphile @Dataphil3]

@trq212

[Andrew Ng @AndrewYNg]

Impressive inference speed from Inception Labs’ diffusion LLMs. Diffusion LLMs are a fascinating alternative to conventional autoregressive LLMs. Well done @StefanoErmon and team!