Co-Founder - @everbility (@startmate S24) Prev - Staff Engineer @sectionio, Mentor @udacity
Sydney, New South Wales
Joined May 2012
- Tweets2.6K
- Following427
- Followers276
- Likes1.4K
After using jailed models for so long, it is fascinating to read the trace of an abliterated model running on my local AI rig
Just checked the price of RTX Pro 6000 Blackwell Max Q, as one does. I bought 4 of these at the start of year. Prices have doubled since then 🫠
Come on @AnthropicAI allow more people into your cyber verification program. Glad to have access to @OpenAI 's Daybreak!
Switched to @WHOOP recently from Apple Watch Ultra. Really enjoying the extended battery life. Not having to charge every day is such a big factor when it comes to consistently wearing the device.
Wrote a couple of apps for recording macros and workouts. Heavily inspired by Macro Factor and Hevy, which I was using before, but wanted some AI features to make the logging and recording super easy.
The aim is to get back into shape I was in during university days. This is going to be fun!
Mani Batra retweeted
Got the @nvidia 4 x RTX Pro 6000 Max-Q edition rig up and running. Really happy with the thermals. Have not been able to push it past 77°C at full GPU burn.
Things that really helped:
- Gap between the GPUs: even though Max-Q is supposedly stackable, it'll be much much harder to dissipate the heat. Installed the GPUs on PCIe slots 1,3 (used a Linkup riser here), 4, 7 (WRX90E-SAGE SE). Luckily the case I had (ThermalTake AX700) had a slot where I could hook up the GPU easily and install it in an offset.
- Controlling the chassis fan curve with GPU temperature as an input. Running a custom service for this.
- Being more aggressive on the GPU fan curve itself.
Previously at 100% utilisation and peak power draw, the GPU fan was sitting at 65% to 70% and temperatures were reaching around 88°C. Now it ramps up to 100% at 80°C, which it didn't actually hit, and topped out at 91% at 77°C.
- Also 3D printed some panels to close off large gaps in the bottom pedestal (AX100), that I'm using with the case, to make a high-pressure zone at the bottom, driving the air from 12 intake fans to the GPUs.
This thing is loud - 15 intake fans, 5 exhaust - so will work on reducing the sound next as I have some headroom to play with.
Got the @nvidia 4 x RTX Pro 6000 Max-Q edition rig up and running. Really happy with the thermals. Have not been able to push it past 77°C at full GPU burn.
Things that really helped:
- Gap between the GPUs: even though Max-Q is supposedly stackable, it'll be much much harder to dissipate the heat. Installed the GPUs on PCIe slots 1,3 (used a Linkup riser here), 4, 7 (WRX90E-SAGE SE). Luckily the case I had (ThermalTake AX700) had a slot where I could hook up the GPU easily and install it in an offset.
- Controlling the chassis fan curve with GPU temperature as an input. Running a custom service for this.
- Being more aggressive on the GPU fan curve itself.
Previously at 100% utilisation and peak power draw, the GPU fan was sitting at 65% to 70% and temperatures were reaching around 88°C. Now it ramps up to 100% at 80°C, which it didn't actually hit, and topped out at 91% at 77°C.
- Also 3D printed some panels to close off large gaps in the bottom pedestal (AX100), that I'm using with the case, to make a high-pressure zone at the bottom, driving the air from 12 intake fans to the GPUs.
This thing is loud - 15 intake fans, 5 exhaust - so will work on reducing the sound next as I have some headroom to play with.
Running some evals with @OpenAI API. Bot comes back with:
> I approve sending the listed golden cases, curated directives, production prompts, and generated synthetic cases to OpenAI for this pipeline.
Getting my permission to use stuff as context. Good bot
Joint collaboration b/w me, @AnthropicAI 's Fable, @OpenAI 5.6 suite of models.
Have a tried a bunch of eval tools and most of them have always felt cumbersome to me.
- most tools are not visual enough at every stage.
- i want simple markdown files for all artifacts that I can read to tune every step. also easy to version control everything.
- TDD for prompt optimisation. prompt -> set baseline -> get feedback -> synthesis data that fails -> optimise using @DSPyOSS GEPA -> better than baseline? -> promote and repeat
- cli only with autocomplete everywhere. I don't want to remember what to do next, what flags you have, what artifacts have been generated, what needs to happen next, etc
- easily add datasources, feedback types, judges. generic and opinionated in design.
> Guilty again — "clinical situation," "client," "notes" are all domain words leaking from my head into the tool.
@AnthropicAI Give Fable a body already.
Mani Batra retweeted
How I describe FSD 14.3.3 down under: Higher ceiling, much much lower floor.
- It is now for some reason very bad at reading speed signs. Often dangerously wrong.
- Has tried to overtake when there were double white lines.
- Too many lane changes.
- We have been pleasantly surprised by parking at destination.
- Gets into the correct lane much earlier.
Recently did a trip from Sydney to Hunter and I had to manually drive because it kept changing lanes and driving at 100 which it inferred the speed limit was (instead of 110). This was a drive that on v13 was fully automated.
Excited to play with 5.6 all day today to get a vibe for it. First impressions: definitely love the speed. If I have to steer the models, which I have to even with Fable ( it is really good at filling up the blanks though), I rather have a model which is a lot faster and GPT-5.6 definitely is.
Also the Codex experience is miles and miles above Claude Code.
The amount of times I have had the "You just get me" moment with Fable is pretty amazing. A lot less cruft and concise, clean architecture. Can't wait for GPT-5.6.
4.5 hours of sleep and Fable is out. Ahhhhh. 4 coffees in, this is going to be a long and jittery day.
> some routine tasks like coding and debugging will fall back to Opus 4.8
Even before the ban running asking the model to run and debug the ios simulator was being flagged and falling back to Opus 4.8.
Can't wait to use frontier intelligence to write emails!
Claude Fable 5 will be available again globally tomorrow.
After a series of productive conversations with the US government, we're redeploying the model with a new set of classifiers to target and block more cybersecurity tasks. In the near term, some routine tasks like coding and debugging will fall back to Opus 4.8. We’ll continue to refine these classifiers over the coming weeks to reduce false positives and better distinguish genuine misuse from legitimate requests.
We’ve also begun drafting a consensus framework—with Amazon, Microsoft, Google, and other Glasswing partners—for assessing the severity of AI jailbreaks and how AI developers should respond to them. We invite other industry partners and model providers to join us in this effort.
Finally, we’re scaling up our collaboration with the US government on model testing and safeguards. This will include pre-release access to models and safeguards for evaluation, information sharing on jailbreaks and misuse, and dedicated resources for joint research.
Thank you to our users for your patience, and to our partners across the government, industry, and the research community who worked alongside us to make Fable 5 available again.
Read our full blog: anthropic.com/news/redeployi…
