Google Releases Rebuilt Gemini 3.5 Pro, Moonshot AI Debuts 2.8T Kimi K3, and AI Uncovers Massive Cancer Research Paper Mill Fraud

Three stories from July 17, 2026 reveal AI's expanding reach — from model architecture to the integrity of science itself. Google DeepMind's Gemini 3.5 Pro is the first major frontier release where a lab scrapped completed base weights and restarted pretraining from scratch due to recursive tool-calling failures — signalling that agentic reliability is now a hard requirement, not a post-launch tuning problem. Moonshot AI's Kimi K3 (2.8T parameters, 896 experts, 16 active per query, Kimi Delta Attention achieving 2.5× efficiency vs K2) is the largest open-weight model ever released and the first to match frontier closed models in agentic reasoning benchmarks at $3/M input tokens. And a Queensland University of Technology (QUT) BERT classifier scanned 2.6 million cancer research papers published 1999–2024, flagging 250,000+ as paper mill frauds — rising from <1% of published papers in 2000 to over 16% in 2022 — with 91% classification accuracy validated against known retracted papers.

---

🤖 Gemini 3.5 Pro — Why Google Restarted Pretraining from Scratch

The Tool-Calling Failure That Forced the Rebuild

Why recursive tool-calling is hard: Standard text generation (completing a prompt) is a single-pass operation. Recursive tool-calling is a loop:

Gemini's initial base model had structural issues at steps 3 and 5 — when receiving tool outputs and using them as inputs to subsequent tool calls, the model incorrectly parsed the schema of intermediate results, causing cascading hallucinations in deep tool chains.

Why Google restarted instead of patching: Post-training fixes (RLHF, SFT on corrected examples) can mitigate surface-level hallucinations but cannot fix structural architectural issues in how the model's attention layers attend to intermediate tool outputs. The problem was diagnosed as originating in the base pretraining weights — not fixable by fine-tuning.

Timeline of the rebuild:

The precedent: Google chose an 8-week delay over shipping a broken agentic model. This is the first documented case of a frontier lab scrapping completed pretraining weights due to tool-use failures — indicating the industry now treats agentic reliability as a launch blocker, not a post-launch problem.

Gemini 3.5 Pro Specifications

Technical specs:

The "Deep Think" mode — what it does: Deep Think is Gemini 3.5 Pro's extended reasoning mode (similar to OpenAI's o3 thinking mode):

• Allocates additional test-time compute before generating the final response
• The model runs internal hypothesis generation and testing (not visible to user, but increases response latency by ~3–8 seconds)
• Trade-off: ~3× slower → ~15–20% more accurate on complex reasoning, coding, and agentic benchmarks

---

🌐 Kimi K3 — 2.8T MoE, 896 Experts, Open Weights

MoE Architecture at 2.8T Scale

Mixture-of-Experts (MoE) — why it enables this scale: In a standard dense model, every parameter is active for every forward pass:

• 2.8T dense model: all 2.8T parameters fire for every token → inference requires moving 2.8T parameters through memory → infeasible

In an MoE model, the model has N expert sub-networks, but only K experts activate per token:

• Kimi K3: 896 total experts, 16 active per query
• Active parameters per forward pass: ~60B (comparable to a dense 60B model)
• But knowledge capacity: encoded across all 896 experts ≈ 2.8T parameters worth of knowledge

Kimi K3 vs dense model economics:

Kimi Delta Attention (KDA) — the efficiency innovation:

Kimi K3 benchmarks vs proprietary frontier:

Kimi K3 is within benchmark noise of the best closed proprietary models — for the first time, open weights = frontier quality.

Pricing and availability:

The MIT license + weights public = any developer can download and deploy Kimi K3 with full fine-tuning access.

---

🔬 BERT Paper Mill Audit — 250K Fraudulent Cancer Papers

The Academic Paper Mill Industrial Complex

What paper mills are and how they work:

Why cancer research is particularly vulnerable:

• High-volume field (2.6M papers 1999–2024 = ~100,000+ papers/year)
• Complex experimental methods (cell cultures, animal models, molecular assays) that are difficult for peer reviewers to verify without repeating the experiments
• High pressure to publish in competitive research environments (China, Iran, Egypt — countries with high identified paper mill activity per the BMJ study)

The BERT classifier — technical approach:

Scale of the findings — the 25-year fraud trajectory:

Impact on oncology research: Papers flagged as paper mill productions have been cited in subsequent research — meaning fraudulent results have propagated into meta-analyses, clinical trial designs, and drug development decisions. If a drug was developed based on a fabricated pre-clinical result, it may have reached Phase I/II trials on a false premise, wasting billions in R&D and potentially exposing patients to ineffective treatments.

Publisher adoption of the BERT classifier: Following the BMJ publication, multiple major scientific publishers announced integration of the BERT classifier into editorial pipelines:

• Elsevier: Integrating as a mandatory pre-review screening step for journals with historically high retraction rates
• Springer Nature: Testing in oncology journal submissions
• Wiley: Evaluating for cell biology + cancer biology journals

The classifier does not reject papers — it flags them for mandatory enhanced human review (second independent peer reviewer + raw data request from authors).

---

📌 The Bottom Line

• gemini-3-5-pro-rebuilt-deep-think-2m-context: First documented case of frontier lab scrapping completed pretraining weights due to agentic reliability failure: recursive tool-calling cascade (A's misinterpretation propagates through B→C→... hallucinatory loop); structural attention-layer issue = not patchable by RLHF; 8-week delay over launching broken agent; specs: 2M context (2× prev), Deep Think mode (3× slower, 15-20% more accurate), 3× nesting tool-call reliability improvement; GPQA Diamond ~80%.
• kimi-k3-2-8t-moe-896-experts-kda: MoE economics: 896 experts, 16 active per query → ~60B active parameters (dense-60B inference cost) but 2.8T knowledge capacity; Kimi Delta Attention: linear attention for long-range + standard for local = 2.5× KV cache memory reduction; benchmarks within noise of best closed models (GPQA Diamond 79.4% vs GPT-5.5 ~78%); MIT license, weights public July 27, 2026; API $3/M input (vs GPT-5.5 $12/M); self-hosted ~$0.20/M.
• cancer-paper-mill-bert-250k-fraud: QUT BERT classifier: trained on retracted (confirmed mill) + genuine papers → detects textual fingerprints (boilerplate, template reuse, stylistic inconsistency), statistical anomalies (impossible p-values), metadata cross-correlation (same author/institution multi-journal same-day pattern), similarity clustering; 2.6M papers scanned, 250K+ flagged (91% accuracy against known retractions); fraud trajectory: <1% (1999-2005) → >16% (2022 peak) — 1 in 6 published cancer papers in 2022 fraudulent; fabricated pre-clinical results cited in meta-analyses + clinical trial designs = potential drug R&D decisions built on false premises; Elsevier/Springer Nature/Wiley integrating as mandatory pre-review screening.

---

📬 Stay Updated

Get the best of AI & technology delivered to your inbox every week. Subscribe to our free newsletter →

---

Disclosure: This post contains affiliate links. If you purchase through our links, we earn a small commission at no extra cost to you. We only recommend products we believe in.