Wanted to introduce Patient Glue, it's an all-in-one SMS platform for healthcare that integrated directly into your EHR.

Currently have an up to 3 month pilot for you to try it for free!

Unlike other service offerings, Patient Glue is made to be budget friendly with a flat-rate per number of seats and usage pricing model.

I added in some AI integrations which are configurable and some automated workflows to make the scheduling process easier.

Built this because I had worked at a smaller practice and the clinic literally couldn't afford those other options.

Check it out yourself, open for any and all feedback!

Remember, happy patients stick around!

Comments URL: https://news.ycombinator.com/item?id=45262520

Points: 1

# Comments: 0

The Idea in a nutshell:

"Hallucinations" aren't indicative of bad training, but per-token semantic ambiguity. By accounting for that ambiguity before prompting for a determinate response we can increase the reliability of the output.

Two‑Step Contextual Enrichment (TSCE) is an experiment probing whether a high‑temperature “forced hallucination”, used as part of the system prompt in a second low temp pass, can reduce end-result hallucinations and tighten output variance in LLMs.

What I noticed:

In >4000 automated tests across GPT‑4o, GPT‑3.5‑turbo and Llama‑3, TSCE lifted task‑pass rates by 24 – 44 pp with < 0.5 s extra latency.

All logs & raw JSON are public for anyone who wants to replicate (or debunk) the findings.

Would love to hear from anyone doing something similar, I know other multi-pass prompting techniques exist but I think this is somewhat different.

Primarily because in the first step we purposefully instruct the LLM to not directly reference or respond to the user, building upon ideas like adversarial prompting.

I posted an early version of this paper but since then have run about 3100 additional tests using other models outside of GPT-3.5-turbo and Llama-3-8B, and updated the paper to reflect that.

Code MIT, paper CC-BY-4.0.

Comments URL: https://news.ycombinator.com/item?id=44294446

Points: 2

# Comments: 0

Comments URL: https://news.ycombinator.com/item?id=44077404

Points: 5

# Comments: 0

Given the demonstrated risk of information leakage from embeddings, have you explored any methods for hardening, obfuscating, or 'watermarking' embedding spaces to resist universal translation and inversion?

Unrelated, but this is exactly why I've been spending time building my AI framework (TSCE). The idea is to leverage these open-weight LLMs, typically smaller and accessible, to achieve accuracy and reliability comparable to larger models. It doesn't necessarily make the models "smarter" (like retraining or fine-tuning might), but it empowers everyday users to build reliable agentic workflows or AI tools from multiple smaller LLM instances. Check it out: https://github.com/AutomationOptimization/tsce_demo

If you reran today you’d see the same improved pass rate I’m getting now. That’s the downside of benchmarking against latest model names; behaviour changes quietly unless you pin to a dated snapshot.

For bigger, noisier prompts (or on GPT-3.5-turbo, which hasn’t changed) TSCE still gives a solid uplift, so the framework’s value stands. Appreciate you checking it out!

+30pp uplift when using GPT-35-turbo on a mix of 300 tasks.

Free open framework, check the repo try it yourself

https://github.com/AutomationOptimization/tsce_demo

I tested this another 300 times with gpt-4.1 to remove those obtrusive "em-dashes" everyone hates. Tested a single-pass baseline vs TSCE, same exact instructions and prompt "Remove the em-dashes from my linkedin post. . .".

Out of the 300 tests, baseline failed to remove the em-dashes 149/300 times. TSCE failed to remove the em-dashes 18/300 times.

It works, all the data as well as the entire script used for testing is in the repo.

Two deterministic forward-passes.

1. The model is asked to emit a hyperdimensional anchor (HDA) under high temperature. 2. The same model is then asked to answer while that anchor is prepended to the original prompt.

No retries, no human-readable scratch-pad, no fine-tune.

---

2. What a hyper-dimensional anchor is

Opaque token sequence that the network writes for itself.

Notation: • X = full system + user prompt • A = anchor tokens • Y = final answer

Phase 1 samples `A ~ pθ(A | X)` Phase 2 samples `Y ~ pθ(Y | X,A)`

Because A is now a latent variable observed at inference time:

`H(Y | X,A) ≤ H(Y | X)` (entropy can only go down) and, empirically, E\[H] drops ≈ 6× on GPT-3.5-turbo.

Think of it as the network manufacturing an internal coordinate frame, then constraining its second pass to that frame.

---

3. Why the anchor helps (intuition, not hype)

4 096-D embeddings can store far more semantic nuances than any single “chain-of-thought” token stream. The anchor is generated under the same system policy that will govern the answer, so policy constraints are rehearsed privately before the model speaks. Lower conditional entropy means fewer high-probability “wrong” beams, so a single low-temperature decode often suffices.

---

4. Numbers (mixed math + calendar + formatting pack)

GPT-3.5-turbo – accuracy 49 % → 79 % (N = 300). GPT-4.1 – em-dash violation 50 % → 6 % (N = 300). Llama-3 8 B – accuracy 69 % → 76 % with anchor alone, 85 % when anchor precedes chain-of-thought (N = 100). Token overhead: 1.3 – 1.9× (two calls). One Self-Refine loop already costs ≥ 3×.

Diagnostic plots (entropy bars, KL-per-position, cosine-distance violin) are in the repo if you like pictures → `figures/`.

---

5. Why this isn’t “just another prompt trick”

The anchor never appears in the user-visible text. Gains replicate on two vendor families (OpenAI GPT and open-weights Llama) and on both reasoning and policy-adherence tasks. Visible chain-of-thought actually loses accuracy on 8 B models unless the anchor comes first; the mechanism changes internal computation, not surface wording.

---

6. Try it yourself

pip install tsce python -m tsce_demo "Rewrite this sentence without any em-dashes — can you?"

Repo (MIT) with benchmark harness, plots, and raw JSONL in Title!

---

7. Questions I’d love feedback on

Optimal anchor length vs. model size (64 tokens seems enough for < 10 B). Behaviour on Mixtral, Phi-3, Claude, Gemini — please post numbers. Red-team attempts: can you poison the anchor in Phase 1 and make the answer leak?

---

Comments URL: https://news.ycombinator.com/item?id=43899325

Points: 3

# Comments: 1

I’d like to share TSCE – a Python library that gives small models the ability to follow instructions with above GPT-4-like precision using a two-step architecture. It’s designed for developers who want reliable, high-quality results without the need for fine-tuning large models.

TL;DR TSCE uses a two-step approach to improve the reliability of language models. This ensures more accurate and coherent outputs, especially for tasks that require rule adherence or detailed instructions.

---

Why You Should Care

* No Fine-Tuning: You don't need to fine-tune models. TSCE works with existing models like GPT-3.5 and GPT-4. * Improved Reliability: By decoupling the thinking process (Phase 1) and the answering process (Phase 2), you get more consistent, on-topic results. * Works with OpenAI and Azure: TSCE works out of the box with both OpenAI and Azure OpenAI APIs, so it’s ready to plug into your existing workflows. * Minimal Overhead: The additional latency is small (~2 API calls), but the payoff is huge in terms of clarity and accuracy.

Example

from tsce import TSCEChat

tsce = TSCEChat(model="gpt-3.5-turbo") prompt = "Provide a brief explanation of the benefits of solar energy and how it impacts the environment. Include an example of a renewable energy source other than solar energy."

reply = tsce(prompt)

print("Answer (Phase 2):", reply.content) # The final response generated after Phase 2 print("Anchor (Phase 1):", reply.anchor) # The structured anchor generated in Phase 1

Output Example:

- *Anchor (Phase 1)*: `solar::harvest~light{clean}—echo::sustain[green]` ###END###`

- *Answer (Phase 2)*: `"Solar energy offers numerous benefits, particularly in terms of environmental impact. Here are some key points:

1. *Clean Energy Source*: Solar energy is a clean and renewable source of power. It generates electricity without emitting greenhouse gases or pollutants, which helps reduce the overall carbon footprint.

2. *Reduction in Fossil Fuel Dependence*: By harnessing solar energy, we can decrease our reliance on fossil fuels, which are finite and contribute significantly to environmental pollution and climate change.

3. *Sustainable and Renewable*: Solar energy is abundant and inexhaustible. As long as the sun shines, we can harness its power, making it a sustainable energy solution for the long term.

4. *Positive Environmental Impact*: Utilizing solar energy reduces air and water pollution, minimizes habitat destruction, and decreases the strain on natural resources. This contributes to a healthier ecosystem and a more balanced natural environment.

*Example of Another Renewable Energy Source*: Wind energy is another excellent example of a renewable energy source. Wind turbines convert the kinetic energy from wind into electrical power. Like solar energy, wind energy is clean, sustainable, and helps reduce greenhouse gas emissions. Wind farms can be established on land or offshore, providing versatile options for generating renewable energy."`

---

Installation

``` pip install tsce ```

Customizing System Prompts

You can easily customize the system instructions for the final pass by passing in a custom *`final_prefix`*:

``` tsce = TSCEChat(final_prefix="You are a helpful assistant, please respond in 3 sentences.") reply = tsce("Explain the concept of recursion.") print(reply.content) ```

---

If you have any questions or want to try it out, feel free to comment here or head over to the repo.

[GitHub Repo](https://github.com/AutomationOptimization/tsce_demo) [GDrive: Read the paper, See the proof](https://tinyurl.com/3xswpzbb)

Looking forward to hearing what you think!

Comments URL: https://news.ycombinator.com/item?id=43788904

Points: 3

# Comments: 0