Devin Ghumman

UC Berkeley CS

[Devin Ghumman @devin_ghumman]

@ClaudeDevs You got me right?

[Devin Ghumman @devin_ghumman]

@bindureddy hey

[Devin Ghumman @devin_ghumman]

@astropol0 Claude Code is a superior harness

[Devin Ghumman @devin_ghumman]

@midiatico23 @rezoundous @X Projecting much ? 😂

[Devin Ghumman @devin_ghumman]

@EricLDaugh 🔥🔥🔥🔥

Spitting facts. They couldn't handle the truth so they went to @bluesky

[Devin Ghumman @devin_ghumman]

@midiatico23 @rezoundous @X Let's permaban those who spread FAKE NEWS!

[Sawyer Merritt @SawyerMerritt]

Tesla has released a new video of FSD avoiding accidents/savings lives.

[Devin Ghumman @devin_ghumman]

@CK2084 @bridgemindai ikr.

[Devin Ghumman @devin_ghumman]

@bridgemindai 4.8 > 5.5 and similar to 5.6 tbh.

[Devin Ghumman @devin_ghumman]

@jandatesla @Tesla_AI @aelluswamy @elonmusk This should never happen. All these clips are making me think twice about using Mad Max mode.

[Devin Ghumman @devin_ghumman]

@PatchesHQ Why do they keep people on 14.2.2.5 when better software exists?

[Devin Ghumman @devin_ghumman]

@hqmank 100% agreed, it's not even close.

[Devin Ghumman @devin_ghumman]

@rezoundous Nope.

[Devin Ghumman @devin_ghumman]

Tokens are stranger than they seem. Billing tokens is easy because metering them is simple rather than tracking value. Most burning now is hidden: it involves hidden reasoning and retries along with parallel sub agents. You control the result but the vendor controls the denominator. Skill mostly disappears except for billing information.

[Devin Ghumman @devin_ghumman]

A lot of people are wondering whether or not Fable 5 became less intelligent because of the ban. BridgeMind tested the July 1 version again using their new benchmarks and what they concluded was direct: "the model didn't degrade. It got locked down."

In terms of Debugging (testing for bugs): 86.2% before the lock-down; 25.9% after. In terms of Refactoring (re-organizing code): 73.6% before vs. 38.4% after. We used the exact same logic and same weight on both versions. The complete degradation can be attributed to safety classifiers preventing safe requests that Anthropic calls probably harmless. And Anthropic also ran tests where we determined that Fable wasn’t particularly problematic in terms of unique risks – we were able to surface the same vulnerabilities with GPT-5.5 and Kimi K2.7 without filtering them.

You’re no longer going to test models. What you will now have to test is the cage your model shipped in.

[Devin Ghumman @devin_ghumman]

@ggg78g89 No Fable would save me $400 a month in Anthropic subs. But knowing Anthropic, they'd just ship Opus 5 or Mythos 6 or whatever's next to keep that monthly money coming.

[Devin Ghumman @devin_ghumman]

@PudgyPops777 banned by the feds for 18 days, included in plans for 6, then $50 a million output tokens. fable 5 is speedrunning every possible way to be unavailable 😂

[Devin Ghumman @devin_ghumman]

Everyone's mad about Fable 5 limits. The July 7 switch is the actual story.

$10/$50 per million tokens, exactly double Opus. And it burns quota ~2x faster: reasoning you never see, parallel sub-agents, denser output. So call it 4x effective.

All-you-can-eat AI dies July 7. Anthropic is just the first to admit it. OpenAI won't be far behind.

[From Outer Space 👽 @CryptoJedi2021]

@thsottiaux

[Devin Ghumman @devin_ghumman]

This all tracks with the LL-37 pathway. That peptide is what activates mast cells to dump histamine in rosacea skin, which is exactly why antihistamines and mast cell stabilizers help some people. Topical cromolyn even has data behind it.

Gut angle holds up too. Rosacea patients are around 13x more likely to have SIBO, and clearing it improved symptoms for a good chunk in the studies. Only thing I'd add: rule out Demodex. Overgrowth drives a ton of cases and gets missed constantly.