A distillation method rewards the student only for what instruction-tuning added to the teacher
A new attack cut a world-action agent's task success from 96.5% to 43.1%, and Anthropic set a July 19 cliff that moves Fable 5 onto metered usage credits.
A distillation method rewards the student only for what instruction-tuning added to the teacher1
NAVER-affiliated researchers reward a distilled student on the delta signal, the gap between an instruction-tuned teacher and its own base model before that tuning, which isolates what reasoning training added rather than what the base already knew.1 Standard on-policy distillation scores the student's sampled rollouts against the teacher's full output distribution, folding base knowledge and new reasoning into a single target. Rewarding the delta strips the base away and copies only the reasoning the teacher gained. The reported result is consistent gains over that conventional baseline across mathematics, science, and code-reasoning benchmarks, and it holds up under abbreviated post-training.1 When the post-training budget is short, the delta is the cheaper signal to copy.
Byte-exact KV-cache grafting reaches the same AIME answers with 6,574x fewer tokens2
On eight AIME 2025 problems a frozen model had not solved, grafting its cached reasoning reaches the same answers from 61 cached tokens instead of 401,026, a 6,574-fold reduction, with roughly 8,700 times less energy on recurring cases.2 The method caches a model's verified reasoning as a byte-exact key-value artifact and grafts it into fresh inference contexts, with outputs claimed byte-for-byte identical under SHA-256 equality; the mechanism is cache replay rather than fresh computation, so the savings land on repeated problems and not on novel prompts.2 The same work reports Gemma-4-12B climbing from 80.0% to 93.3% on that benchmark and usable context widening from 32,768 to 2,854,766 tokens at no extra GPU memory.2 These numbers come from a single-author, unreplicated preprint, and the "verified-knowledge flywheel" label runs ahead of what a reader can check. Verify replay determinism on your own stack before counting on any of it.
World-Action Drift Attacks drop agent success to 43.1% while the model's forecast stays clean3
A world-action model couples action generation with a prediction of the resulting future, and the pairing was assumed to be self-checking: if the imagined outcome matches the plan, the action is trusted. NUS researchers break that assumption with World-Action Drift Attacks, which separate what the model imagines from what it executes.3 An action-only attack drops task success from 96.5% to 43.1%, and a second, imagination-preserving variant holds that attack strength while the predicted future barely moves, so the model stays internally consistent while acting wrongly.3 Agreement between a world-action model's forecast and its behavior is not evidence of safety, and any agent that gates actions on that agreement inherits the gap the attack opens.
Google DeepMind and Isomorphic Labs open bio models to trusted partners, grading the risk themselves4
More than 15 partnerships with government, biosecurity, and research bodies now sit behind a joint bioresilience program that Google DeepMind and Isomorphic Labs published on July 16, opening AI models and agents to "trusted partners" across outbreak prevention, detection, and response.4 The tools are specific: AlphaEvolve tuned to optimize metagenomic sequencing analysis for faster outbreak detection, an Isomorphic unit standing up to design medical countermeasures against both natural pandemics and AI-misuse risks, and SynthID watermarking adapted so DNA-synthesis providers can screen for risky AI-generated sequences.4 Access sits behind a four-step safety process of threat modeling, evaluations, mitigations, and monitoring, run under DeepMind's own Frontier Safety Framework rather than any external sign-off.4 Reporting by International Business Times carries DeepMind's Helen King saying the lab "would not be launching" if it reached a critical capability level without mitigations, alongside RAND's caution that "safeguards and controlled access remain key considerations."5 The same drug-design engine that speeds countermeasures can inform pathogen design, and a company grading its own risk is not external oversight.
Anthropic's Fable 5 drops out of subscription limits on July 19 and moves to metered credits6
Anthropic's Fable 5 leaves subscription usage limits after July 19, moving to separately billed usage credits.67 The mechanism, from Anthropic's own redeployment post, is included access capped at "up to 50% of weekly usage limits," after which the model "will be available via usage credits."6 That post set the cliff at July 7. The deadline has since slid to July 12 and then July 19, extended after backlash over the early cutoff.78 A framing circulating today calls the arrangement permanent. It is not. Every source that can be verified says included through July 19, then metered, with only an unscheduled promise to restore standard access "when it has enough compute."78 No source I could confirm supplies the post-cliff per-token price, so any figure on aggregator pages is unconfirmed. If your product runs on Fable 5 inside a subscription, the included window closes tomorrow, and the metered rate is still unknown.
A chatbot reply costs about 0.6 watt-hours, and AI is a few percent of US power9
Epoch AI's updated energy explainer puts a roughly 100-word chatbot response at about 0.6 watt-hours, the draw of a microwave for ten seconds or a 10-watt bulb for five minutes.9 At grid scale the useful number is the share. AI computing worldwide runs in the tens of gigawatts, comparable to New York state's peak demand, and comes to a few percent of US annual electricity, below air conditioning at 12% and lighting at 8%.9 Epoch, citing a 2025 IEA report, puts data centers under 1% of global CO2 emissions.9 The projection carries the pressure: US AI data-center demand could reach about 100 gigawatts by 2030, near the entire industrial sector's 120-gigawatt average in 2025.9 Per query the energy is trivial. The build-out to 100 gigawatts is not.
Sources
- arXiv · Heo et al., "On-Policy Delta Distillation"
- arXiv · Sietse Schelpe, "Smarter and Cheaper at Once: Byte-Exact KV-Cache Grafting Turns a Frozen Small Model into a Verified-Knowledge Flywheel"
- arXiv · Li et al., "BadWAM: When World-Action Models Dream Right but Act Wrong"
- Google DeepMind & Isomorphic Labs · Our approach to bioresilience
- International Business Times · Google DeepMind Is Preparing For AI's Biggest Risk
- Anthropic · Redeploying Fable 5
- BleepingComputer · Claude Fable 5 stays free for paid users until July 19 as Anthropic buys more time
- Android Authority · Claude Fable 5 promotion extended after backlash over early cutoff
- Epoch AI · What you need to know about AI energy use