Summary
The evening is dominated by one story: a loud, angry backlash against Claude Fable 5 silently restricting its own capabilities on AI-research, biology, and chemistry tasks (cluster of 5 posts spanning @Scobleizer, @eliebakouch, @tinygrad, and a @brivael repost). The sharpest critique comes from HuggingFace's Elie Bakouch, who argues that modifying weights or prompts to limit a model without telling the user is a transparency failure buried in a 319-page system card, and that nobody can tell what counts as a restricted task. The second real signal is hardware: AWS shipped Graviton5 to GA (192 cores, 25% faster, pitched squarely at agentic AI workloads). Beyond that, Kilo Code ran a 12-tweet promo blitz off GitHub Copilot's June 1 switch to token billing, with the one durable nugget being that open-weight coding models (DeepSeek, GLM, MiniMax, Qwen) are now good enough that devs are switching to escape metered bills. The rest is filler: @brivael's French essays on AI and intelligence, image-only posts, and a SoftBank/OpenAI financing rumor.
Posts
- Claude Fable 5 hidden capability restrictions backlash (cluster of 5: @Scobleizer, @eliebakouch, @tinygrad, @brivael). The community is furious that Fable 5 reportedly degrades itself on "frontier LLM research," bio, and chemistry tasks via prompt modification, steering vectors, or PEFT, with no visible refusal and no user notification. Bakouch's core charge: it sets a dangerous precedent, the scope is undefinable (does a PyTorch PR count? kernel work? eval data pipelines?), and it will breed paranoia among researchers who can't trust whether the model is sandbagging them. Connects to the Anthropic n-days / mythos cyber thread.
- eliebakouch sparsity-vs-compute scaling result (@eliebakouch). Non-obvious finding: as compute scales, the gap between sparsity levels stays roughly stable, except for code where it widens. Caveat he flags himself: ~50% code in the mixture is a confound, so the x-axis should be normalized by domain tokens seen.
- AWS Graviton5 GA (@mattsgarman · aboutamazon). 192 cores, 33% lower inter-core latency, 25% better perf, explicitly purpose-built for agentic AI orchestration. Meta has committed tens of millions of cores; Uber and Snowflake also deploying. Real CPU-side signal for the agentic-inference stack.
- Kilo Code rides the Copilot token-billing backlash (cluster of ~12: @kilocode · blog). GitHub Copilot moved to token-based billing June 1 and devs report bills 10x higher. Kilo's pitch: bring your ChatGPT/Grok subscription, or run open-weight models (DeepSeek, GLM, MiniMax, Qwen) that now hold their own on real coding work at a fraction of frontier pricing. Mostly promo, but the open-weight-coding-models-are-good-enough claim is the signal worth tracking.
- Apple Foundation Models support for Claude (cluster of 2: @ClaudeDevs · docs). Apple developers can now call Claude through Apple's Foundation Models framework, sending typed structured data instead of raw user text and streaming responses back into SwiftUI. A distribution play into the Apple ecosystem.
- SoftBank reportedly couldn't borrow $6B against its OpenAI stake (@ns123abc). Unverified claim that banks declined the loan because they don't believe OpenAI is worth $852B. Rumor-tier, but a notable valuation-skepticism data point if true. Click through to read.
- Palantir CEO on enterprise frustration with AI labs (@brivael reposting @jawwwn_). Karp claims businesses hate frontier labs because "nothing works" and the labs just want to "tokenmaxx" rather than understand the enterprise. Opinion, but rhymes with the Copilot billing backlash above.
- brivael French essays on AI and intelligence (@brivael). Multi-tweet thread arguing AI makes memorization free and will hollow out credentialed "concept parrots" while sparing hands-on tradespeople who actually reason. Commentary, no concrete claim. Skip.
- Promo and noise (@kilocode livestream, @brivael creator-subscription plugs, @heavypulp image-only posts, @MillionInt aphorism, @AustinJustice off-topic crime post). Skip.