• Higgsfield Blog (rss) · 5 gundur sessiz

2026 in LLMs (so far)

• <p>On Friday I gave the closing keynote at the <a href="https://www.wearedevelopers.com/world-congress-north-america">WeAreDevelopers World Congress North America</a> in San Jose. I tied together the key trends from the past year into a chronological exploration of everything that happened in 2026.

TypeSafe AI's Jev launches a new 'System One' model category for fast, cheap structured decisions—triggering rapid open-source alternatives like Laya and Stanford/Nvidia's CLM-8B.

• TypeSafe AI released Jev on Sept 15: a non-generative model returning typed decisions (e.g. 'billing' vs text) in 70-500ms at $0.042/M input tokens with no output-token cost.
• Jev claims 40-200x lower latency and 444.6x cost reduction vs frontier LLMs for classification/routing/scoring tasks.
• Open-source alternative Laya (421M params, ModernBERT-based) offers a Jev-compatible API and ~33ms inference on a Tesla T4.
• Stanford/Nvidia's CLM-8B uses contrastive embeddings instead of classification heads, running up to 9x faster than Jev in zero-shot tests but sometimes with lower accuracy (CLM: 81.6%/87.6% vs Jev: 71.1%/83.1% on coding-judge benchmarks).
• Emerging pattern: 'Jev-as-judge' replaces expensive LLM-as-judge—generative models produce candidates, cheap decision models rank/route them, escalating only low-confidence cases to LLMs or humans.
• Trained via RLCD (Reinforcement Learning for Calibrated Decisions) so confidence scores are meaningful—0.52 vs 0.998 can trigger different handling logic.

Crusoe's new Serverless Fine-Tuning platform publishes full benchmarks and costs, showing small fine-tuned models can beat giant ones at a fraction of the price.

• Crusoe's Serverless Fine-Tuning is now GA, covering 19 open models from 2B to 750B parameters, each tested with a public, reproducible benchmark harness.
• Fine-tuning GLM 5.2 (753B params) cost just $180 and boosted intent classification by 19pp and multi-turn conversation scores by 61pp.
• A fine-tuned Qwen3.5-2B outperformed Qwen3-235B on the Banking77 customer-service benchmark at roughly 1/15th the training cost.
• Pricing is fully transparent and tiered: $0.40 per million tokens under 16B params up to $10 above 300B, no sales calls required.
• Crusoe openly reported regressions (Qwen3.5-9B and Gemma-4-31B got worse on a theorem-proving benchmark after fine-tuning), flagging likely pre-training contamination instead of hiding it.
• Fine-tuned models can deploy straight to dedicated endpoints via Self-Serve Deployments starting at $5.50/hour on Nvidia H100s.

CrowdStrike's product chief says AI agents don't just attack faster, they chain tactics into novel attack patterns that break human-reasoning-based defenses.

• AJ Shipley (CrowdStrike CPO) argues AI agents 'reason differently' than humans, combining known TTPs into unfamiliar attack chains that 20-30 year old security architectures weren't built to detect
• Entry-level SOC jobs are shifting up-stack: AI now handles tier-1 alert triage and initial evidence-gathering, so 'entry-level' will mean tier-2/tier-3 work instead
• Shipley frames this as a productivity multiplier, not headcount replacement — same analyst count expected to do 10x-100x more work with AI tooling
• CrowdStrike's proposed fix for AI non-determinism: ensemble of models 'voting' on outcomes plus input-controlling harnesses, similar to old multi-engine malware sandboxing
• 3-5 year bull case: 'red models' auto-discover vulnerabilities, 'blue models' auto-remediate them across endpoint/cloud/SaaS/network before exploitation
• Sponsor segment (Coder) pushes concept of a dedicated 'AI Operating Layer' sitting outside the app/data/compute stack for governance and visibility

Kākāpō Party

• <p><strong>Tool:</strong> <a href="https://tools.simonwillison.net/kakapo-party">Kākāpō Party</a></p> <p>I presented a closing keynote for the <a href="https://www.wearedevelopers.com/world-congress-north-america">WeAreDevelopers World Congress North America</a> yesterday. As <a href="https://simonw