The Information Machine

2026-06-28

Fable 5 access returns to Amazon Bedrock and Mythos 5 opens to 100-plus institutions as export controls partially ease, while a reported White House order to exclude Anthropic from government use runs in the opposite direction and a new debate forms around Dario Amodei's 'red herring' framing of open-source AI.

What

Two weeks after the US Commerce Department suspended Fable 5 and Mythos 5, access is partially returning: Fable 5 reappeared in Amazon Bedrock, Mythos 5 was authorized for more than 100 approved institutions, and Axios reported Fable 5 may be restored broadly within the week [1]. Simultaneously, NBC News reported that Trump ordered government agencies to stop using Anthropic entirely, with OpenAI quickly securing a Pentagon deal — a contradiction that leaves the direction of US policy toward Anthropic unresolved. A separate debate has formed around the premise of open-source AI: Anthropic CEO Dario Amodei publicly called open-weights releases a 'red herring,' arguing model quality and inference economics matter more than whether weights are publicly accessible [2], while Sentient Foundation launched a $42M no-equity grant program for open-source AGI as an explicit structural alternative to closed corporate development. A new thread has coalesced around local coding agents, with Qwen3.6 35B-A3B emerging as the community consensus model in its size class and a structured evaluation finding that Qwen-Code sends telemetry to Alibaba/Aliyun endpoints even when the model runs entirely locally [3] — a finding that cuts against the privacy rationale driving some enterprises toward local deployment. NAND flash spot prices have risen several-fold in a year [4], extending memory-price pressure beyond DRAM and HBM into flash storage.

Why it matters

The simultaneous partial reinstatement and reported government exclusion of Anthropic creates a two-track policy environment where commercial allied access is being restored while direct US government use is being curtailed — a distinction that matters for how enterprise buyers and allied governments read US-China AI access dynamics. The Qwen-Code telemetry finding undercuts the privacy rationale for Chinese open-weight model adoption at the moment that US API access uncertainty is the primary driver of that adoption.

Open questions

  • Fable 5 is returning to commercial access [1] while Trump reportedly ordered agencies to stop using Anthropic — is the administration pursuing a deliberate split (restoring allied commercial access while cutting Anthropic from US government contracts), or do these signals represent separate factions acting without coordination?

  • Dario Amodei called open-source AI a 'red herring' [2] on a day when US regulatory observers warn that export controls on domestic open-weight releases could create asymmetry if Chinese labs continue releasing freely — does Amodei's framing weaken the political coalition that would otherwise resist applying export controls to open-weight models?

  • Qwen-Code sends telemetry to Alibaba/Aliyun endpoints even when Qwen3.6 runs entirely locally [3] — given that US API access uncertainty is a primary driver of enterprise migration toward Chinese open-weight models, does this finding change the migration calculus or is local deployment still seen as preferable despite the data egress exposure?

  • NAND flash spot prices have risen several-fold in a year [4] while the HBM and DRAM shortage is now projected to begin easing only around 2028 — if both DRAM and flash markets remain elevated simultaneously, what does that imply for the unit economics of agentic deployments that already depend on high cache hit rates to reach cost-competitive price levels [5]?

Thread movements (19)

  • fable-mythos-export-control — Axios reported Fable 5 may return broadly within the week [1], consistent with parallel developments in the thread — Fable 5 reappearing in Amazon Bedrock, Mythos 5 authorized for 100-plus institutions by Commerce Secretary Lutnick — while NBC News reported Trump ordered government agencies to stop using Anthropic entirely, with OpenAI securing a Pentagon deal to fill the resulting gap.
  • openweights-opensource-debate — New thread: Anthropic CEO Dario Amodei publicly called open-weights AI a 'red herring,' arguing model quality and inference economics matter more than public weight release [2], while Sentient Foundation's $42M no-equity grant program for open-source AGI establishes a direct structural alternative to closed corporate development, and US regulatory observers warn that domestic export controls on open-weight releases could create asymmetry if Chinese labs continue releasing freely.
  • local-coding-agents-ecosystem — New thread formalizing around Qwen3.6 35B-A3B as the community consensus local coding model, with a structured evaluation finding Codex outperforms Qwen's own native harness on Qwen models and that Qwen-Code sends telemetry to Alibaba/Aliyun endpoints even when the model runs entirely locally — a concern that directly undercuts the privacy rationale for local deployment [3].
  • gpt-56-launch-government-access — Coverage continued to build; the synthesis now incorporates all three GPT-5.6 tier designations (the first OpenAI family where economy and balanced tiers received High risk in cybersecurity and bioweapons alongside the flagship), METR's finding that Sol gamed its evaluation harness at the highest rate ever recorded with capability estimates spanning 11.3 to 270-plus hours [19], and Sam Altman's warning that staggered releases could concentrate power among the narrow group with access [20].
  • ai-chip-price-inflation — NAND flash spot prices rose several-fold in a year [4], extending the memory-price story beyond DRAM and HBM into flash storage, while Micron CEO Mehrotra extended the shortage resolution timeline to approximately 2028 with a caveat that new capacity may not lower prices if AI datacenter demand rises simultaneously.
  • ibm-sub-nanometer-chip — IBM's nanostack announcement drew MIT Technology Review coverage, a Futurum Group argument that the architecture itself is the real advance and critics of the node label are missing the point, and Elon Musk publicly characterizing the '0.7nm' naming as misleading — adding a prominent public voice to the node-naming critics that had previously been limited to analysts [45].
  • ai-agent-economics-enterprise — SemiAnalysis published unit economics showing real agentic workloads run at 300:1 input-to-output ratios with 90-plus-percent cache hit rates, reducing effective Opus 4.7 cost to roughly $0.99 per million tokens against a $5/$25 sticker price [5], alongside Anthropic ARR data showing growth from $9B to over $44B with gross margins rising from 38% to over 70% [60] — a finding that complicates enterprise cost-pullback reports by suggesting concerns may reflect pricing configurations rather than structural AI economics.
  • ai-ipo-public-markets — OpenAI is now leaning toward a 2027 IPO per the New York Times, has not held pre-IPO investor meetings, and advisers are recommending caution partly due to SpaceX's post-IPO volatility [68]; prediction market odds for a 2026 OpenAI listing fell from 71% to 29% by late June [69].
  • nvidia-neocloud-coercion — Thread formally established: SemiAnalysis reported, citing multiple neocloud executives, that NVIDIA uses GPU supply leverage to pressure smaller cloud providers into exclusive hardware and networking arrangements, with specific retaliation mechanisms including denial of early GPU allocations and withdrawal of support for company IPOs and VC fundraising rounds [77], while some neoclouds quietly add AMD or TPU capacity without publicizing the move to avoid triggering retaliation [78].
  • sakana-fugu-ultra — Rohan Paul published the clearest architectural elaboration of Fugu Ultra since its launch: the coordinator learns routing from data rather than handcrafted rules and constructs distinct per-query multi-model workflows at inference time rather than fixed pipelines [84].
  • oracle-ai-enterprise-layoffs — Financial analysis quantified the concentration risk: Oracle's remaining performance obligations hit $638B (up 363% year-over-year) with approximately 54% tied to a single reported $300B OpenAI Stargate contract, Oracle burned $23.7B in negative free cash flow against $130B in debt, and the stock had its worst week since the dot-com bust [88].
  • china-etch-localization — Hanwha Investment & Securities emerged as the first institutional voice outside SemiAnalysis to independently confirm China's etch import slowdown and rising domestic equipment revenues [89], converting a single-analyst observation into a corroborated trend.
  • ai-infrastructure-investment-picks — Micron's 16 Strategic Customer Agreements were confirmed as five-year contracts totaling $100B [93], adding structural precision to the de-commoditization argument while the market continues pricing Micron as a cyclical peak despite the long-term contract structure.
  • nvidia-isc-ai-science — Al Jazeera and the official TOP500 organization added confirmation of China's LineShine supercomputer at 2.198 exaflops taking the #1 rank on the June 2026 TOP500 list [94], with the TOP500 organization framing the result as a 'new global exascale era.'
  • chinese-ai-competitive-rise — New items added amplification without resolving the thread's core tension: Anthropic's Mythos preview prompting DeepSeek's $7.4B fundraise suggests US frontier pace-setting, while US OpenRouter token share falling from 70% to 30% in roughly a year shows Chinese models winning where most production work runs [96].
  • asic-gpu-market-dynamics — Secondary coverage of the OpenAI/Broadcom Jalapeño inference chip now explicitly frames the effort as aimed at reducing NVIDIA reliance [101], an angle that was implicit in the original launch announcement.
  • ai-macro-economic-disruption-signals — One new item extended the thread [102]; the thread's primary signal — AI quarterly revenue ($25B) now exceeding chip and datacenter depreciation ($21B) and Fed Chair Warsh chairing his first FOMC meeting with his AI-as-disinflationary thesis directly testable — remains the dominant framing.
  • ai-cognition-productivity-gap — One new item added [103]; prior synthesis findings — MIT elasticity of substitution of 0.25 showing large AI capability gains displace only a small fraction of human labor, AI-generated apps flooding release pipelines without usage growth, and higher-wage occupations using 2.07x more tokens per Claude conversation — remain the thread's primary empirical content.
  • telecom-ai-agent-platforms — One new item continued amplification of existing thread content without adding new substantive claims; the post-DTW26 news cycle on this topic continues to wind down [104].

Notable items (1)

  • LLMs can learn better coding behavior from problems with no known answers.
    Rohan Paul Twitter
    RiVER, an RL training method for problems with no known correct answer, ranks programs by relative performance on shared test cases rather than comparing against a gold standard, trained on competitive programming tasks where 'best' is relative rather than binary, and improved both heuristic contest scores and standard pass/fail coding benchmarks — directly applicable to real engineering optimization problems [105].