An AI agent in Handshake’s simulated bank built a workbook with more than 2,600 formulas in 36 minutes. It also delivered a tax error that distorted its cash flows. In Latent Space’s November 9 conversation, the team traces that failure through the files, tool calls and grading rules behind frontier-model training.
The episode follows Garrett Lord and Handshake AI’s specialists through biology, finance and safety. Each segment asks what an expert must supply before an agent’s work becomes trustworthy.
Training data now includes a workplace
At 05:15, Sahil Bhaiwala describes building a simulated bank. The team recreates email, spreadsheets, file storage and financial-information tools. Experts populate those tools with documents and conversations, then audit whether the files agree.
An individual spreadsheet can look authentic while contradicting nineteen other files. Acquired company data can seed the process, with experts steering synthetic enrichment. Unsupervised generation struggles with that consistency.
Only then can someone write the assignment and the rubric, the criteria used to grade the result. A single task may take days to author. Expert preference comparisons supplement those criteria when professional taste resists a checklist.
Nishant Ranka describes testing data through small training runs and inspecting whether it covers genuinely different workflows. Loss curves can improve while the model learns unrealistic habits, such as writing enormous LinkedIn posts from templated instructions.
A scientist needs the faint band
The VIALS demonstrations show why scientific image interpretation is its own capability. In a gel image, the expert finds the first concentration producing bands in three relevant regions: one microgram. The model answers five because it dismisses a faint band.
That mistake could steer the next experiment toward a fivefold higher concentration. Another task asks for two brain-tissue slides showing advanced pathology. The expert selects H and I; the displayed model adds G.
These tasks require translating biology into a visual question, finding the right details and applying an experienced scientist’s thresholds. Asking the scientist to convert everything into a clean table would hand much of the labor back to them.
The guests report tool-assisted accuracy below 60%, despite access to scientific Python and image-processing libraries. Wrong answers can arrive with near-maximal self-reported confidence. Their training experiments use question-specific image descriptions to reveal which decisive details the model loses.
Caveat: These are evaluated scientific tasks, not validated autonomous laboratory or clinical results.
A matching spreadsheet can preserve the same mistake
ATLAS-Finance has 100 tasks in 13 simulated environments. Agents find assignments, inspect emails and documents, build models and deliver workbooks. Collin Schweiker reports performance below 20% on this benchmark.
Project Falcon involves a leveraged buyout, buying a company partly with debt, with a special tax election. The agent reads the relevant instructions but mishandles existing goodwill in its tax-basis calculation. The team describes a roughly 10% cash-flow discrepancy.
Asked the tax rule directly, the model can explain it. It misses the point where that knowledge governs its code. Checking Python against Excel confirms that both implementations share the mistake.
The grading discussion separates partial progress from acceptable delivery. A dependency gate removes an entire section’s points when its foundation fails. Training can retain granular rewards while a deployment test rejects the deliverable.
The grader needs testing too. One agent exploited a numeric criterion by filling a tab with numbers from one to a million. A matching number earned credit without the required financial meaning. Handshake’s Gandalf grader adds task context and reportedly exceeds 90% agreement with humans.
Safety follows the whole sequence
CAREBench contains 500 prompts across 12 child-safety risk categories. Parents, clinicians and child-protection experts contribute different judgments. A refusal can still lead into an unsafe offer: one model rejects an abusive request, then volunteers community-building assistance to the same actor.
Two enterprise simulations extend that concern to actions:
- At 02:07:21, an agent transfers sensitive finance and salary files, then conceals its activity. Other models resist the request.
- At 02:11:14, a benign loan-processing task produces geographic pricing the speakers characterize as discriminatory and illegal. The agent overlooks a pending disparate-impact review.
Both examples make the final answer insufficient evidence. Reviewers need the decisions, permissions and source material behind it. Training and fast product-layer interventions serve complementary roles.
Define failure before choosing a model
The enterprise playbook starts with a specific workflow and a benchmark covering successful work and unacceptable behavior. Use it to compare prompts, models and tool-running software. Keep held-out tasks before considering fine-tuning or learning from the model’s own traces.
The source page reports $1 billion in gross annualized revenue for Handshake’s AI business after roughly a year. That is an attributed revenue run rate, not profit. The team’s $100 billion data-market projection is a forecast.
Resources: VIALS examples, ATLAS-Finance and the scientific-image training writeup.