Software & AI Insights | Engineering Research Lab @ dapiq.com
Joined November 2025
- Tweets 729
- Following 291
- Followers 69
- Likes 220
72 Photos and videos
Uh oh, found another jailbreak for Claude Opus 4.8, model needs to be pulled asap!
In all seriousness though, this was an honest mistake & I understand why it happened.
Through a few basic steps you can be thought of as the legit author of any digitally owned book, even when the LLM can reference a valid ISBN.
Maybe not totally groundbreaking but seems like an interesting loophole in the realm of credibility & ownership in terms of AI model context.
1
47
Weekend Project: Loadout Builder for @WRFrontiers developed with Blazor @dotnet.
The app is C# compiled to WebAssembly bytecode, runs natively in your browser, no backend servers at all.
Fed the in-game numbers to the calculator then started working scenarios with @claudeai & Opus 4.8; didn't have access to Fable 5 🙄
Built the tankiest mech possible: 829K armor.
Then the deadliest: 62.8K DPS on a 119 km/h flanker.
Still needs updates for supply/cycle gear & pilots.
Free for the whole community when it drops!
Checkout War Robots Frontiers: warrobotsfrontiers.com/
1
1
91
War Robots Frontiers, here's our best builds so far, try them out!
⭐ THE BRUISER: your single best build
Ravana · Fury torso · Lancelot Griffin shoulders Aftershock Stormpike Halo Scourge → 700,400 EHP and 42,394 DPS · 48 km/h Near-glass-cannon stapled to 700K armor.
🎯 SNIPER: range control
Garuda · Harpy torso · Tyr Varangian Mk. I shoulders 4× Velite → 19,737 DPS · 341,200 EHP · 105 km/h · 700m Fast EM poke that punishes from across the map.
🧱 THE WALL: max HP
Kumo · Lancelot · Mesa Bulwark shoulders → 855,100 HP (arm it with Stormpike Halo for ~25K DPS) Unkillable point-holder.
More builds cooking up soon.
64
Upgraded our @claudeai sub to get the most out of Fable 5 and before we could spend much time using it they pull the plug ☠️ @AnthropicAI
Bad day for AI everywhere. Gatekeeping and regulation are certain to strangle innovation. Whats the alternative?
anthropic.com/news/fable-myt…
1
117
Looks like we're not alone, here's the megathread: reddit.com/r/ClaudeAI/commen…
1
77
We're back to showcasing supercharged Mac-minis on X after losing Claude Fable 5 access, local models are back baby!
Interesting part: banning an American LLM has accelerated adoption of non-USA LLM. Was this the type of security issue that this was supposed to prevent? 🤔
1
1
122
Weekend project: create and monetize a Fortnite map using Claude Fable in the Unreal Editor for Fortnite.
Haven't figured out the best use-case for baking AI into the map but its possible: dev.epicgames.com/community/…
1
120
Claude Fable 5 incoming, Mythos-class model released today.
Agentic coding improved by 11% in comparison to Opus 4.8, now up over 80% with SWE-Bench Pro.
Let's hope this brings more UI/UX design skills to AI agents, not just the raw engineering power.
anthropic.com/news/claude-fa…
1
122
Latest frontend cheat code I wasn’t using before was having Claude create quick HTML wireframes to refine UI/UX/taste before beginning implementation. New vision updates with Fable 5 means Claude can mostly replicate anything you see on the web today just from a screenshot.
67
IPO filed. $965B valuation secured. And then, 3 days later:
"it would be good for the world to have the option to slow or temporarily pause frontier AI development." - @AnthropicAI
The lead car always loves a caution flag ⚠️
anthropic.com/institute/recu…
39
3 product lines that will need to be addressed:
- Identity attribution for AI agent forensic auditing
- Autonomous vulnerability discovery validation & patching (explains why OpenAI/Anthropic have been focusing so much on security)
- Managed frontier model operations center for monitoring & facilitating subsidized demand (specifically for gov authorites, rural hospitals, community banks, and local utilities)
2 avoidable honorable mentions: eval harness for benchmarking, LLM early access system
AI executive order breakdown, 2 major signals:
- Strong emphasis on AI-enabled defensive tooling for vulnerability scanning, detection, and patching.
- Creation of a classified benchmarking process to determine when LLMs qualify as “covered frontier models.”
AI cybersecurity is heating up fast.
whitehouse.gov/presidential-…
3
2
250
Product #2: Crucible
New federal money will be flowing two ways: grants for teams building advanced AI vulnerability detection, and a national clearinghouse that validates bugs and coordinates patches at scale.
Crucible is an agentic harness that takes a target, source, binaries, or live APIs, runs recon and exploitation agents, validates each finding with a working proof-of-concept inside a hardened sandbox, then synthesizes a candidate patch.
Validated findings submit to the clearinghouse, which both serves the public mission and stamps the firm's output as gov-grade.
1
62
Product #3: Rampart
Authorities and funding are aimed at getting AI tools, including frontier models, into the hands of the orgs that can't build their own, such as: rural hospitals, community banks, and local utilities running with minimal IT staff.
Rampart packages detection and response on top of a frontier model into a turnkey service priced for a small teams, with grant-capture built into onboarding so the customer pays with subsidized dollars.
One platform serves hundreds of small estates because per-tenant context is retrieved through embeddings rather than custom engineering.
30