NeuroGenesys.
Research Desk • Category

Machine Learning

14 articles on machine learning from the NeuroGenesys Research Desk, newest first. Every claim is linked to its source.

  1. A tall branching tree made of fine cyan light lines standing on a wet night-time salt flat under the Milky Way, with an amber beam climbing its trunk and lighting a single violet branch tip while the others stay dim, all reflected in the hexagonal salt crust
    • 5 min read

    Agents still have the right skill. They stop picking it as the library grows

    Two papers from October and November 2026 measure what happens to LLM agents as their skill libraries grow into the hundreds and thousands. The skills keep working; the agents stop choosing them. That is the failure our forgetting-as-selection results predict, and both papers fix it without touching the skills.

  2. A row of equal violet light rods standing along a dark horizon at dusk, with one amber column rising far above them into a starry sky until its top fades out
    • 7 min read

    Superintelligence has many definitions and no test. Here is what a test would need.

    Labs now say superintelligence is in sight, but their definitions compare AI with different humans, and none comes with a way to check. Benchmarks saturate, ceilings are set by label errors, and the best long-task metric stops working at 16 hours.

  3. Small violet points of light scattered across a dark plain at night, with tall amber beams rising from a few of them and large stretches of the plain left dark
    • 8 min read

    Where AI already beats the best humans in 2026, and where it still doesn't

    AI now matches or beats top humans on olympiad maths, contest programming and most reasoning benchmarks, and has disproved a famous conjecture. It still fails on long reliable work, learning after deployment and the physical world, and many of the strongest claims are graded by the companies themselves.

  4. A line of violet light points running low across a dark salt flat that bends upward into an amber arc and fades into the night sky before it is clear where it ends
    • 8 min read

    The case for a 2027 intelligence explosion now rests on evidence the public can't see

    Forecasters moved their dates for automated AI research back to 2027 this week, on the strength of METR's finding that Anthropic is improving about twice as fast as a year ago. Most of the evidence behind that finding is redacted, and the gains may be limited to tasks that grade themselves.

  5. Five faint constellations over a dark desert, one lit amber by a narrow beam from the horizon while the other four stay dim violet
    • 5 min read

    Penalising wrong answers only works if someone else is grading

    Experts that never saw a task still vote on it with near-certainty. We ran ten pre-registered fixes. Penalties inside the expert failed; a selector that ignores expert confidence reached 0.926 on Split-MNIST.

  6. A web of faint violet points of light above a night-time salt flat, with a few new amber points streaking in and the violet points nearest them glowing brighter
    • 6 min read

    To stop a model forgetting, rehearse what it knew. New context won't do it.

    We tested whether bridging knowledge added alongside new facts protects old ones. It doesn't, on a toy model or on Qwen 1.5B. Re-expressed old knowledge does, and outdated facts need pruning.

  7. A tall branching tree made of fine cyan light lines standing on a wet night-time salt flat under the Milky Way, with an amber beam climbing its trunk and lighting a single violet branch tip while the others stay dim, all reflected in the hexagonal salt crust
    • 6 min read

    Investors are paying for agents that learn on the job. The hard part is knowing what they forgot.

    Trajectory just raised $45M at a $300M valuation to make deployed models keep learning. But on CL-Bench, the best system gained only 25.4% over a stateless baseline, and plain long context beat dedicated memory.

  8. A glowing prism of light lines on a night-time salt flat splitting one amber beam into hundreds of cyan and violet threads, only a few of which pass through a bright amber ring-shaped gate, under the Milky Way
    • 6 min read

    Tokens keep getting cheaper and AI bills keep binding. Count verified tasks instead.

    Apple cut some engineers' daily AI budget from $300 to $200 and Revolut moved coding to self-hosted GLM at about $0.30 per million tokens. Agentic workloads multiply tokens faster than prices fall. The unit that matters is cost per verified task.

  9. A stream of violet and amber points of light passing through a glowing arched gate above a night-time salt flat into a calm violet lattice, with a few red points held back
    • 6 min read

    What a model lets into long-term memory matters more than how it learns

    In our experiments, a per-example trust gate cut a poisoning attack from 94% to 4%, where a sleep phase added only 3.6 points. The leverage is in what gets consolidated.

  10. A restless violet neural lattice and a small, regular amber lattice floating above a night-time salt flat, joined by thin beams of light
    • 6 min read

    Can a guardian chip stop a learning model from forgetting? Only partly.

    A protected core model on its own silicon can watch a continually learning model for drift. The hardware is ready; the judgment about what to keep is not.

  11. A glowing violet neural lattice floating above a night-time salt flat, with a small cluster of amber nodes growing from one edge
    • 5 min read

    To keep what a model knows, control what it is allowed to move

    Meta cut forgetting from 89% to 11% by updating 500 memory slots. Across continual learning, the methods that hold up limit what moves and keep old data in play.

  12. A glowing violet lattice floating above a night-time salt flat, bent where a single amber beam touches it from above
    • 4 min read

    A model rewrote another model's weights. Read what it took away.

    An MIT agent edited 0.016% of a small Qwen model and its math scores jumped. Across weight editing in 2026, subtraction works and addition still lags.

  13. One beam of light splitting into many near-identical parallel rays above a salt flat, with a few rays of different colours crossing them
    • 3 min read

    More thinking isn't more evidence: test-time compute meets correlated error.

    Sampling a reasoning model sixteen times does not give you sixteen opinions. New preprints put numbers on an old statistical warning.

  14. Separate streams of light rising from the horizon and braiding into one luminous river in the sky
    • 7 min read

    Decoders that generalize: the case for brain foundation models.

    Speech neuroprostheses crossed the usefulness threshold by borrowing the modern ML playbook — and the next leap is models that transfer across sessions, subjects, and tasks.

Leadership

A word from the founder — on why the brain is the blueprint, and why the next decade of intelligence will be built here.

"The brain is the universe's only proof that general intelligence is possible. We treat it as the blueprint."

From cortical interfaces to photonic silicon, every program at NeuroGenesys exists to answer one question: how does intelligence arise — and how do we build it safely.

Will Schulz

Will Schulz

Founder, NeuroGenesys • AW3 Technology